Head to head: Anthropic: Claude Opus 5.5 vs Meta: Muse Spark 1.3

Anthropic: Claude Opus 5.5: # Arcwell's Relay Uses AI to Feed Answers to Customer Support Agents Arcwell is betting that the answer to most customer questions already exists somewhere inside a company, and that support agents lose too much of their day hunting for it. The company's new platform, Relay, uses artificial intelligence to read customer support conversations and suggest responses to the agents handling them. It can also flag customers who may be about to cancel and generate summaries. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. Arcwell says early testing supports that pitch. Twelve companies used Relay over six months, and agents working with it closed tickets 14% faster, according to the company. Arcwell did not release customer-retention results, which leaves open whether the cancellation warnings actually helped those companies hold on to customers. Relay costs $65 per agent each month, so the bill grows with the size of a support team. To do its job, Relay needs access to a company's support transcripts and internal help documents, material many businesses will want to handle carefully. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period for their data. Those assurances matter because Relay's suggestions are only as good as the information it can see. A tool built to surface what a company already knows has to be allowed to look through it first. For prospective buyers, the evaluation comes down to what Arcwell has shared and what it hasn't. The company has offered a 14% speedup in ticket resolution, drawn from its own six-month test, and a per-seat price. It has not yet published evidence that Relay keeps customers from leaving, even though spotting those customers is one of the platform's core features.

Meta: Muse Spark 1.3: # Arcwell Bets on AI Assistants, Not Replacements, With Relay Launch Arcwell is launching Relay, an artificial-intelligence platform that reads customer-support conversations as they happen and suggests ready answers for human agents to send. Rather than automating replies outright, the system sits alongside agents. It surfaces likely answers during live chats and tickets, flags customers who show signs they may cancel, and writes summaries once a case is resolved. For now, the emphasis is on assistance rather than automation. The pitch is time. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said. Relay is designed to pull that information forward, reducing toggling between tabs and internal wikis when queues are growing and patience is thin. Arcwell tested Relay with 12 companies over six months. Agents using the software closed tickets 14% faster, the company said. It did not release customer-retention results, leaving open the question of whether faster resolutions translated into fewer cancellations — the harder metric for the churn-prediction feature to prove. Relay will cost $65 per agent per month, a per-seat model familiar to support leaders already paying for help-desk software by the seat. That access comes with tradeoffs. To make relevant suggestions, Relay requires access to support transcripts and internal help documents, the raw material it uses to understand past resolutions and company policy. For buyers sensitive about what leaves their systems, Arcwell says customer data is encrypted and is not used to train shared models, meaning one company’s tickets won’t shape answers for another. Companies can choose a 30-day or 90-day retention period for stored data. How customers react to AI-assisted support remains to be seen, but Arcwell is framing Relay as help for overwhelmed teams rather than a replacement for them. If the time savings from the pilot hold up more broadly in busy real-world queues, the company argues, agents can spend less effort hunting for answers and summarizing calls, and more on the conversations that keep customers from leaving.

Anthropic: Claude Opus 5.5: # Startup Spotlight: LedgerLark Wants AI to Handle the Tedious Side of Post-Trade Work Most talk about AI in finance focuses on the front office: trading signals, research summaries, client chatbots. LedgerLark is looking at what happens afterward, in the back-office work of confirming that a trade recorded in one system matches the same trade everywhere else. The startup builds AI agents for post-trade operations at small investment firms. It is an unglamorous corner of finance, but mistakes there can be expensive and slow to untangle. ## A small error, a large cleanup Co-founder Mina Batra spent seven years in operations at a regional brokerage. She began exploring the idea after watching a six-person team spend three days resolving a single trade mismatch. The cause was a copied account number. That incident shows the problem LedgerLark is targeting. A small data-entry error can spread across confirmations, custodian records and internal books. Finding where the records diverge often means people manually comparing documents line by line. Co-founder Owen Carr brings experience from the other end of the industry. He built reconciliation systems at a large bank. The pairing combines someone who has done the work by hand with someone who has automated it at scale. Their bet is that smaller firms need something in between. ## What the product does LedgerLark's agents compare three sources of truth: - Trade confirmations - Custodian records - A firm's internal ledgers When the agents find a discrepancy, they propose a correction. A human then approves or rejects it. The software does not execute trades or move money. That limit is a meaningful design choice. It reduces the potential damage if the system gets something wrong, and it may make the product easier for cautious compliance teams to accept. But it also means the product's value depends on the quality of its proposals. If staff end up checking each suggestion as carefully as they would have investigated the original exception, the time savings shrink. ## Early traction LedgerLark has five paying customers, with annual contracts ranging from $36,000 to $110,000. It has raised a $3.5 million seed round led by Motive Row. The company says one customer reduced its unresolved daily exceptions from 70 to 18. If that holds up, it is a notable improvement. Still, it comes from a single customer and has not been independently verified. Several questions remain open: - Over what period was the reduction measured? - How does the customer define an exception? - Were the remaining exceptions the hardest cases? - How much did the team's own workflow changes contribute? Five customers is a real start, but it is too early to say how well the results generalize. ## Competition and friction LedgerLark is not entering an empty market. Established reconciliation vendors already sell to financial firms, and large banks run their own internal software. Many existing tools rely on matching rules that are predictable and well understood. LedgerLark will need to show that its agents handle messy, irregular cases better than those rules do, without adding new kinds of risk. Sales could also be slow. Financial institutions may require lengthy security and procurement reviews before adopting new software, especially software that touches records of client trades. For a seed-stage company, long sales cycles can strain both time and cash. ## What to watch Several signals will show whether LedgerLark's approach is working: - Whether additional customers report results similar to the 70-to-18 figure - How often human reviewers accept the agents' proposed corrections - How long it takes to move new firms through security and procurement reviews - Whether incumbent vendors add comparable AI features LedgerLark is tackling a concrete, recognizable problem with a deliberately limited product. The hard part is proving that its agents save enough time to justify switching tools, in an industry that is careful by necessity.

Meta: Muse Spark 1.3: # Startup Spotlight: LedgerLark Takes On Post-Trade Busywork For retail investors, a trade feels finished when an order is filled. Inside small investment firms, that moment often starts a scramble. Someone must match confirmations against custodian records and internal ledgers, chase small inconsistencies, and ensure everything settles correctly. At smaller firms, that work often falls to lean operations teams. LedgerLark is building AI agents focused on that post-trade workload for small investment firms. The idea is not to automate trading itself, but to help operations staff find and resolve the mismatches that can otherwise pile up at the end of the day. ## An operations problem Co-founder Mina Batra spent seven years in operations at a regional brokerage before starting LedgerLark. Co-founder Owen Carr previously built reconciliation systems at a large bank. Batra says the idea began to take shape after watching a six-person team spend three days resolving a single mismatch caused by a copied account number. The error was small, but tracing it across records and agreeing on a fix took sustained effort. It raised the question behind the company: how much operations time goes to routine discrepancies rather than genuinely complex issues? Carr's background is relevant. Large banks often have internal tools and dedicated teams for reconciliation. Small firms generally do not, so similar problems can consume a larger share of staff time. ## What the product does LedgerLark's agents compare three sources — trade confirmations, custodian records, and a firm's internal ledgers — and flag where they diverge. When the system spots a likely cause, it proposes a correction for a person to review and approve. That human approval step matters. The product does not execute trades or move money, a boundary the company describes as deliberate. In practice, staff still decide whether a proposed fix is correct. That may ease concerns about automating financial operations, but the tool's value depends on suggestion quality and how much time review actually saves. ## Early customers and early numbers LedgerLark reports five paying customers, with annual contracts ranging from $36,000 to $110,000. That range suggests the product is being sold as an operations tool rather than core infrastructure, at least for now. One example the company points to: a customer reduced unresolved daily exceptions from 70 to 18 after adopting LedgerLark. That figure comes from the company, reflects one deployment, and has not been independently verified. Still, daily exceptions are a concrete metric for this work, so it will be worth watching whether similar patterns appear elsewhere. The company has raised a $3.5 million seed round led by Motive Row. ## A cautious market with entrenched options LedgerLark is not entering an empty field. Established reconciliation vendors already serve financial firms, and many larger institutions rely on software built in-house over years. For small firms, spreadsheets and manual checks remain common, in part because switching systems carries its own risk. There is also the question of sales cycles. Financial institutions may require lengthy security and procurement reviews before adopting tools that touch ledger data. Even if the agents only suggest corrections, firms will likely want to understand how data is handled and audited. That could slow adoption beyond the first handful of customers. The constructive question is whether small firms will see enough day-to-day relief to justify that evaluation effort — and whether LedgerLark's suggestions prove reliable across different custodians, formats, and edge cases. For now, LedgerLark offers a narrow test of AI agents in finance: not replacing operations staff, but reducing the morning backlog. Whether that leads to broader adoption will depend less on novelty than on sustained accuracy, clear audits, and patience with enterprise sales.

Anthropic: Claude Opus 5.5: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

Meta: Muse Spark 1.3: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

Anthropic: Claude Opus 5.5: # Raincheck Raises $8 Million to Help Cities Prepare for Floods *The Austin startup's software combines drainage maps with rainfall forecasts to help local governments decide where to send maintenance crews before a storm.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. Basin Ventures led the round, with participation from existing investor Long Mile. The new funding brings Raincheck's total raised to $10.5 million. Chief Executive Ana Flores said the software addresses a problem cities often face before a storm: > "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." Raincheck describes its approach in [an overview of its forecasting methodology](https://example.com/raincheck-methodology). The company works with seven cities in Texas and Louisiana. It declined to identify them because some contracts have not been publicly announced. Raincheck's forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Raincheck competes with other flood-planning vendors, including HydroMap and municipal engineering contractors. The company says it will use the new funding to hire eight engineers and expand into Florida.

Meta: Muse Spark 1.3: # Raincheck raises $8 million to help cities get ahead of floods *The Austin startup combines drainage maps with rainfall forecasts to guide pre-storm maintenance* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The round was led by Basin Ventures, with participation from existing investor Long Mile. The company has now raised $10.5 million in total. Raincheck maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. The company outlines [its forecasting methodology](https://example.com/raincheck-methodology). Said CEO Ana Flores: > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

Anthropic: Claude Opus 5.5: **C**

Meta: Muse Spark 1.3: C

Anthropic: Claude Opus 5.5: **6** 1. `#Flood software moves upstream` has no space after `#`, so it breaks the headline rule. 2. `The company is expanding its pilot.` sits in the dek position but is not italicized. 3. `##Details` has no space after `##`, so it breaks the section-heading rule. 4. `Read [the announcement](www.example.com/news).` uses a link without an absolute HTTPS URL. 5. `* The first pilot begins Monday.` starts the list item with an asterisk instead of a hyphen. 6. `<div>The second pilot begins Friday.</div>` is raw HTML.

Meta: Muse Spark 1.3: 6

Anthropic: Claude Opus 5.5: # The Top Two Support AI Models Tied on Quality. One Costs 61% Less. When researchers put five language models through a customer-support test, the most useful result was not which model scored highest. It was that the top two models performed about the same, and one of them costs far less to run. Model North had the highest average score, 84.2. Model Vale was close behind at 83.8. The researchers found that this 0.4-point gap was not statistically significant, so the test cannot reliably tell the two models apart on overall answer quality. Their costs, however, were very different. North was estimated at $1.90 per 1,000 tickets and Vale at $0.74. That makes Vale about 61% cheaper. Vale was also 18% faster. For a support team choosing a model from these results alone, Vale offers about the same quality as North at a lower price and with faster responses. ## How the test worked The researchers collected 200 anonymized customer-support tickets and had each of the five models write one proposed answer per ticket. Three experienced support managers then graded the answers without knowing which model wrote which response. Blind grading matters here because it reduces the chance that a reviewer's expectations about a particular model could influence the scores. Each answer received a combined score based on three criteria: factual correctness, usefulness and tone. This reflects how support managers usually judge real replies. An answer that is accurate but cold, or friendly but wrong, is not a good answer. A combined score also has a drawback. It does not show which of the three criteria caused a difference between models. Two models with similar averages could have gotten there in different ways, such as one being more accurate and the other more polished. The published summary does not split out those parts. ## The wider field: a cost-quality curve Below the top two, the results follow a familiar pattern. Cheaper models generally scored lower. - **Ember** scored 79.1 at $0.31 per 1,000 tickets. - **Kite** scored 76.4 at $0.22. - **Harbor**, the cheapest at $0.08, scored lowest at 71.0. The summary reports a significance test only for the North–Vale comparison, so readers should be careful about small gaps lower in the rankings. The difference between Ember and Kite, for example, may or may not be meaningful. The larger pattern is clearer. Harbor scored about 13 points below North, which suggests that the cheapest option involves a real loss in quality. Whether that loss is acceptable depends on how the model is used. A model that drafts replies for a human agent to review can be weaker than one that sends answers straight to customers. ## Where each leader shines North and Vale were close overall, but they did not perform the same way on every type of ticket. North did best on billing disputes, and Vale did best on technical troubleshooting. The summary does not give the size of these differences, so they should be treated as signals to investigate, not as firm conclusions. Still, the split could affect real buying decisions. A company whose support queue is mostly billing complaints might reasonably test North further, even though it costs more. A software company with many technical tickets has another reason to prefer Vale. ## A shared blind spot All five models struggled with one kind of ticket: those that depended on an image attachment, such as a screenshot of an error message. This weakness appeared across every model regardless of price. Choosing a more expensive model did not solve it. Customers often send screenshots, so teams should plan for these tickets. Possible approaches include sending them to human agents or adding a separate step that interprets the image before the model writes a reply. ## What this benchmark doesn't tell you The test had a narrow scope, which limits how far its results apply. All 200 tickets were in English and came from a single software company. A retailer, a bank or a company serving customers in many languages could see different results. The researchers also did not test agentic actions, tool use or multilingual support. In this test, the models only wrote answers. They did not look up account records, issue refunds or carry out other tasks that modern support systems increasingly automate. A model that writes well may not perform as well when it has to take actions. Finally, each model produced only one answer per ticket. This shows typical performance but not how consistent each model is when given the same ticket more than once. ## The takeaway Within these limits, the main result is clear. Paying more than twice as much for North did not produce a measurably better answer than Vale. Teams evaluating support models should run the same kind of blind test on their own tickets. They should also pay particular attention to their own mix of ticket types and to the screenshot-dependent cases that every model in this test handled poorly.

Meta: Muse Spark 1.3: # When Second Place Wins **Top quality tied — but one model cost less than half as much.** For support leaders shopping for AI, the question is no longer whether language models can draft a decent reply. It's whether paying more gets you a meaningfully better one. A new benchmark suggests that, for straightforward text tickets, the answer is no. Researchers ran five models — North, Vale, Ember, Kite, and Harbor — on 200 real, anonymized tickets from a software company. Each model produced one proposed answer per ticket. Three experienced support managers scored the answers blind, without knowing which system wrote which, on a blend of factual correctness, usefulness, and tone. That method matters: it measures what a manager would actually let go to a frustrated customer, not trivia or vibes. ### A tie at the top On raw averages, North led at 84.2, with Vale just behind at 83.8. That gap is noise. The researchers say the difference was not statistically significant. For practical purposes, they tied. Price and speed were not tied. North cost an estimated $1.90 per 1,000 tickets. Vale cost about $0.74 — less than 40% as much — and answered 18% faster. In support, where thousands of tickets turn into real money and queues build in minutes, that is decisive. Equal quality for far less money and less waiting wins, even from second place. Ticket type added nuance. North was strongest on billing disputes, where precise policy language and empathy count. Vale led on technical troubleshooting, where stepwise accuracy counts. Neither dominated overall, a reminder to weight any benchmark toward your own mix rather than a single headline score. ### What cheap costs you The other three models show the tradeoff for chasing the lowest price. Ember (79.1) and Kite (76.4) clustered several points behind the leaders, while Harbor lagged at 71.0. Their estimated costs were dramatically lower, at $0.31, $0.22, down to just $0.08 per 1,000 tickets. Tempting on a spreadsheet, risky in an inbox. A five- to 13-point deficit on a scale built on correctness and usefulness shows up as missing steps, invented policies, or tone-deaf replies that create escalations — costs that quickly erase inference savings. That doesn't rule budget models out. As supervised assistants for first drafts, triage, or classification with human review, they may still pay. But this test gives no reason to trust them as autonomous drafters. ### Where everyone failed The most important finding wasn't about the winner. When a ticket depended on an image attachment — a screenshot of an error, a photo of a setup — all five models performed poorly. Text-only drafting collapses the moment a customer writes “see attached.” Support inboxes are full of screenshots. Any deployment will need vision capability, smart routing, or a human in the loop for those cases, no matter how high a model scores on plain text. ### What this can't prove Context limits the lesson. All tickets were English-language, from one software company. A retailer, bank, or telecom with different jargon and stakes could look different. Other languages could differ more. The study also tested only one-shot answers. It did not test agentic actions, tool use, account lookups, refunds, or multilingual support — increasingly the real job of support agents. A beautiful writer that can't call the right API is only half an agent. Costs, too, are inference estimates only. They exclude implementation, guardrails, evaluation, and the human oversight you will still need. ### The bottom line Don't buy the leaderboard. Buy the tradeoff. Here, Vale offers effectively North-level quality at a fraction of the cost and latency. Unless your queue skews heavily toward billing nuance where North excelled — and you've validated that edge on your own data — Vale is the rational default. More broadly, the test marks a shift. As leading models converge on quality, operational factors decide: price per thousand, response time, fit to ticket types, and handling of images and actions. Score alone won't tell you what to ship.

Anthropic: Claude Opus 5.5: # Foundry Hall to close after losing its largest sponsor Foundry Hall, the Milwaukee nonprofit that has helped launch more than 200 startups since opening in 2018, will close October 31 after its largest corporate sponsor declined to renew its support. Blank River Bank supplied 38% of the accelerator's annual budget. Without that money, Foundry Hall has enough to operate through October but not enough to run another full program. Six employees will lose their jobs, and the twelve startups now working out of its space must leave. Executive director Camille Ross said the organization had hoped another funder would step in to cover the gap. "We kept waiting for the replacement check that never came," she said. A Blank River Bank spokesperson said the decision reflected a change in the bank's community investment priorities and was unrelated to Foundry Hall's performance. By the accelerator's own measure, that performance was considerable. The 212 startups it has supported collectively employ 1,140 people. Those jobs are spread across companies that, in many cases, began with little more than a founder and an idea. Dev Malik is one of those founders. He said his first meeting with an investor took place inside Foundry Hall. Before that, he worked wherever he could find a seat. "I was pitching from the back table of a coffee shop," he said. For the twelve companies still in the building, the closing raises an immediate practical question: where they will go next. Deputy Mayor Susan Cole said the city is talking with building owners and local universities about space for the current companies. Her statement did not name specific sites or say when arrangements might be made. The search for space addresses only part of what the companies stand to lose. Foundry Hall offered more than desks. It brought early-stage founders together under one roof, close enough to trade advice and, at times, connected them with the investors they needed. Malik's account suggests how much that could matter. A founder working alone from a café and a founder meeting investors in a dedicated space are, in practice, running different kinds of businesses. None of this changes the arithmetic Ross faced. A single sponsor carried more than a third of the budget, and when that sponsor stepped back, no one stepped forward. For a nonprofit, that kind of dependence can go unnoticed for years and then decide everything in a single season. The accelerator's final weeks will be spent winding down. Staff members will look for new work, and the remaining startups will pack up and move. The companies that passed through Foundry Hall earlier will carry on elsewhere. Many of them, like Malik's, can trace their beginnings to a building that will soon be empty. When the doors close at the end of October, the 1,140 jobs will remain. The place where many of them started will not.

Meta: Muse Spark 1.3: # The Last Cohort at Foundry Hall **MILWAUKEE** — After eight years of turning coffee-shop ideas into funded companies, Foundry Hall will close on October 31. The announcement, made to staff and founders this week, ends months of uncertainty about its funding future. The nonprofit startup accelerator, which opened in 2018 in a converted warehouse on Milwaukee's near west side, has supported 212 startups that now collectively employ 1,140 people. By the end of the month, its lights will go out. Its mentorship nights and demo days became a familiar entry point for first-time founders across the region. Executive director Camille Ross said the closure follows the loss of the organization's largest corporate backer. Blank River Bank, which supplied 38% of Foundry Hall's annual budget, declined to renew its sponsorship. With enough cash to operate through October but not to run another full program, the board voted to shut down. “We kept waiting for the replacement check that never came.” The decision will leave a tangible hole. Six employees will lose their jobs, and 12 startups currently working out of the Hall's shared workspace must find new homes by the closing date. Laptops, prototypes, and whiteboards still covered in dry-erase plans will have to be packed into boxes. For many, it was less an office than a daily routine of borrowed advice and overheard progress. For founders like Dev Malik, the loss is harder to measure. He arrived in Milwaukee with little more than a pitch deck and built his logistics company inside the accelerator. He recalled holding his first investor meeting in the building after months of pitching from the back table of a coffee shop. The bank said its withdrawal was not a judgment on the accelerator's work, describing it instead as a shift in its community investment priorities. City officials say they are trying to soften the landing. Deputy Mayor Susan Cole's office is talking with building owners and local universities about short-term space for the displaced companies, though no agreements have been announced. For now, those conversations have offered reassurance more than guarantees. Ross said staff are spending their final weeks helping founders close out milestones, forward mail, and save contact lists built over years of demo days. There will be no final cohort, no graduation. What remains, for now, is a quieter version of the usual bustle — founders finishing what they started, in a place that taught them how to start. Alumni have stopped by to take photos by the lobby mural and exchange numbers, as if preserving something that was never only about desks or Wi-Fi.

Anthropic: Claude Opus 5.5: # HelioDock Launches Night Shift Software for Remote Oversight of Warehouse Robots *The autonomous-robot maker also closed a $24 million Series B, though one beta customer says the product's usefulness depends on connectivity.* HelioDock, which makes autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor their robot fleets remotely. Alongside the release, the company announced a $24 million Series B round led by North Coast Ventures. Night Shift targets a persistent problem in automated warehouses. When a robot stalls or gets stuck, someone often has to be on site to sort it out, even in the middle of the night. HelioDock is betting operators will pay a monthly fee to handle more of those moments from wherever they are. That pitch depends on reliable network connections, and the product enters a market where rivals already sell fleet-management tools. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," HelioDock CEO Mara Velez said. ## What the beta showed Before Thursday's launch, Night Shift spent 14 months in a private beta with nine logistics companies running a combined 310 robots. HelioDock says customers in that program experienced 17% fewer unplanned robot stoppages. The figure has not been independently verified. Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity. During rural LTE outages, he said, it became less useful. That experience reflects how the software behaves when a connection drops. The robots return to their locally configured behavior, and remote control becomes unavailable. The fleet keeps running on the rules set at the facility, but an off-site operator cannot step in until the link returns. For warehouses with dependable networks, the gap may matter little. For sites that rely on cellular service in rural areas, it could determine how much of the overnight problem Velez described the software can solve. ## Funding The Series B brings HelioDock's total funding to $39 million, which means the company had raised $15 million before this round. North Coast Ventures led the new financing. ## Pricing and competition Night Shift will cost $90 per robot per month when it becomes generally available on September 15. At that rate, a 100-robot fleet would cost $9,000 a month, or $108,000 a year. A deployment the size of the entire beta program, with 310 robots, would run about $27,900 a month. Buyers are likely to weigh those costs against the alternative Velez described: keeping staff on call to travel to a facility when a robot stops. Whether the subscription pays for itself will depend on several factors. These include fleet size, how often robots halt unexpectedly and how much troubleshooting can be done without someone on the floor. HelioDock's claimed reduction in stoppages, if it holds up outside the beta, would factor into that calculation. Customers will also have alternatives. Relay Robotics and YardOS both offer competing fleet-management products. That means HelioDock is not introducing remote fleet oversight to the market. Instead, it is competing for operators who can compare its features, reliability and price against existing options. For companies already running HelioDock robots, the comparison may come down to a narrower question than features alone. They will need to decide whether the convenience of remote control justifies $90 per machine each month, given connectivity conditions at each of their sites. Pike's experience suggests that question could have different answers within a single company's network of facilities, depending on where each warehouse is located and how reliable its connection is.

Meta: Muse Spark 1.3: # HelioDock Launches Night Shift for Remote Monitoring of Warehouse Robots *After a 14-month private beta with nine logistics companies and 310 robots, the $90-per-robot software arrives September 15 alongside a $24 million Series B* HelioDock, which makes autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely. The release makes remote supervision a standard paid option for HelioDock customers and sets a price and date for broader use after extended testing. HelioDock also announced a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million. Night Shift will cost $90 per robot per month and become generally available September 15. Chief executive Mara Velez described Night Shift as a way to avoid sending staff to a facility in the middle of the night to address a stopped robot. Night Shift lets operators monitor those robots remotely. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” Velez said. The scenario Velez cited involves a single robot stopped by a pallet in the early hours. Instead of driving across town at two in the morning, a manager could rely on remote monitoring of the fleet. Night Shift is presented as the software for that remote view. Night Shift spent 14 months in a private beta before Thursday’s launch, according to HelioDock. Nine logistics companies took part, operating 310 robots with the remote-monitoring software installed. The company is using that period of operational use as the foundation for general availability. The beta fleet of autonomous warehouse robots provided the day-to-day context for monitoring from outside the facility, rather than a short trial. HelioDock says beta customers experienced 17% fewer unplanned robot stoppages while using Night Shift. The company presents the figure as an outcome of remote monitoring of its autonomous warehouse robots. The figure has not been independently verified. Velez’s example of a robot confused by a pallet illustrates the type of interruption that operators watching remotely would encounter. Aaron Pike of Fulcrum Logistics, one of the beta participants, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. His assessment ties the value of remote monitoring to network conditions. Where connectivity was strong, Pike described effective performance. Where rural LTE service faltered, he said the software offered less utility for watching robot fleets. According to the company, a loss of connectivity changes what Night Shift can do. When connectivity fails, robots return to their locally configured behavior. Remote control becomes unavailable at that point. The locally configured behavior then governs the robots without input from remote operators monitoring the fleet. Operators using Night Shift to monitor fleets remotely would not have remote control during those periods. Night Shift will be offered alongside other options for fleet oversight. Competing fleet-management products are available from Relay Robotics and YardOS, giving logistics companies a choice of suppliers for software to monitor robot fleets. HelioDock’s entry does not arrive in an empty category. Night Shift will cost $90 per robot per month, the company said. The per-robot rate means the monthly amount for any operator depends on the size of the fleet placed on the software. General availability is set for September 15, moving the product beyond the nine-company beta. HelioDock also announced a $24 million Series B led by North Coast Ventures. The round brings the company’s total funding to $39 million.

Anthropic: Claude Opus 5.5: # Cloudnote Debugging Page Exposed Prompt Excerpts From 1,842 User Sessions *The page was open to the public without a login for four hours Monday. The company says it has found no sign that anyone beyond the researcher who reported it viewed the data, but its investigation is not finished.* A debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday, Cloudnote said, leaving fragments of text from 1,842 user sessions open to anyone who reached it. Each excerpt could contain up to 200 characters of a prompt. That limit constrains what was exposed, but people often type sensitive material into prompts, and a short passage can still hold a name, a private detail or confidential work information. Cloudnote said full documents, passwords and payment information were stored separately and did not appear on the page. The 1,842 sessions are a small number compared with Cloudnote's 3.2 million registered users. However, sessions and users are different measures, and a single person can account for more than one session. The company has not said how many individual users were affected. Cloudnote removed the page 19 minutes after a researcher notified it of the problem. The company says it has found no evidence that anyone other than the researcher accessed the page. Because its investigation is still underway, that conclusion is preliminary. Cloudnote has not yet said how the page became public or whether the exposure began earlier than its logs currently show.

Meta: Muse Spark 1.3: # Cloudnote exposed prompt snippets from 1,842 sessions for four hours *Company says it removed an unauthenticated debugging page within minutes of being alerted and has found no evidence of other access, but its investigation is ongoing* Cloudnote, the note-taking app with 3.2 million registered users, left a debugging page exposing excerpts of customer prompts accessible on the open internet without a password for four hours on Monday, the company said. The page displayed snippets of up to 200 characters from 1,842 user sessions, according to logs reviewed by the company. Cloudnote said it took the page down 19 minutes after an outside researcher flagged it, and that passwords, payment information and full documents were stored in separate systems and did not appear on the page. What remains unclear is whether anyone other than the researcher saw or copied the data during the exposure window. Cloudnote said it has so far found no evidence of additional access, but emphasized its investigation is not complete and it cannot yet rule out that possibility. For users, the limited length and temporary exposure narrow — but do not eliminate — the potential risk. Security experts generally advise treating any prompt text as potentially sensitive. Cloudnote has not yet said whether it will notify affected users, leaving open key questions about who was affected and what steps, if any, they should take.

Anthropic: Claude Opus 5.5: # KiteBridge Software Glitch Sends $2.8 Million in Duplicate Reimbursements Ninety-one companies paid out about $2.8 million they never owed this week after a software update at expense-management provider KiteBridge sent 13,400 duplicate reimbursements to their employees. The failure occurred despite the company's earlier promise that its system would prevent exactly this kind of error. By Wednesday evening, KiteBridge said $2.3 million had been recovered. The remaining roughly $500,000 was still pending. The cost of the duplicate payments fell on the employers, which fund reimbursements through the platform. No employees had money withdrawn from their accounts, the company said. ## How a routine update went wrong KiteBridge makes software that companies use to process employee expenses, such as travel and meal costs, and to pay back workers who cover those costs out of pocket. On Tuesday at 8:45 a.m., the company deployed version 4.2 of its software. The problem began with a timeout at the payment processor, the outside service that moves money between bank accounts. A timeout happens when a system sends a request and receives no response within an expected window. Software is commonly designed to retry in that situation, because the original request may simply have been slow or lost. Retrying is safe only if the system can tell that a second attempt is the same payment rather than a new one. KiteBridge's software failed to make that distinction. When the processor timed out, a fallback system, a backup path meant to keep payments flowing when the primary route has trouble, treated the retries as fresh reimbursement requests. Each one was paid. The result was thousands of employees receiving the same reimbursement more than once. The first customer complaint reached KiteBridge at 9:12 a.m., less than half an hour after the update went live. The company suspended automated payments at 10:03 a.m., about 50 minutes after that first complaint and roughly an hour and 20 minutes after the deployment. By then, the duplicate payments had reached $2.8 million. ## A safeguard that didn't hold The episode is likely to draw scrutiny because KiteBridge had told customers it had addressed this risk. An April changelog, the running list of changes a software company publishes with each release, said version 4.0 had introduced "exactly-once payment protection." In payments engineering, "exactly once" describes a guarantee that each transaction is carried out one time and only one time, even if messages are delayed, lost or resent. Tuesday's incident shows that, at least under the conditions created by version 4.2 and the processor timeout, that guarantee did not hold. ## Slow disclosure KiteBridge's public account of the problem also lagged behind events. Its incident page, where the company posts updates on service disruptions, initially described the issue as delayed reimbursements. That description reflected the effect of pausing payments but left out that money had already gone out in error. The page first disclosed the duplicate payments around 4 p.m. Tuesday, six hours after the company had halted automated payments. That gap meant customers checking KiteBridge's official status updates would not have learned from them that their accounts had been used to overpay employees. ## Company response Chief Executive Lena Ortiz apologized for the incident and said KiteBridge would publish a technical report explaining what went wrong. The company has not said when the report will be released, how it plans to recover the remaining funds, or when automated payments will fully resume.

Meta: Muse Spark 1.3: # KiteBridge Glitch Sent $2.8 Million in Duplicate Payments to Workers **91 employers paid staff twice after a failed update; most of the money has been recovered** Ninety-one companies started Tuesday thinking they had a routine payday ahead. By midmorning, they had paid thousands of their employees twice. Expense-management software company KiteBridge said a software update released Tuesday morning malfunctioned, creating 13,400 duplicate reimbursements totaling $2.8 million. The erroneous payments went to workers who had filed legitimate expenses, but the cost fell on their employers, who funded the extra payouts. No money was withdrawn from employees’ accounts. The failure is especially embarrassing because KiteBridge had explicitly promised to prevent it. The company rolled out version 4.2 of its platform on Tuesday morning. Shortly after the release, KiteBridge’s system encountered a timeout while talking to its payment processor — the outside service that actually moves money. In plain terms, KiteBridge sent payment instructions but did not get a timely confirmation that they had arrived. That kind of communications hiccup is common, and payment software is supposed to be built to survive it. When a confirmation doesn’t arrive, the system retries. To avoid paying twice, modern systems tag each payment with a unique ID, so that if the same request shows up again, the system recognizes it as a retry and processes it only once. KiteBridge said its fallback system did the opposite. When retries were triggered by the timeout, it treated each one as a new reimbursement request. The result was a second, identical payment for thousands of expense reports that should have been paid a single time. Finance teams noticed quickly. The first customer complaint arrived within about 30 minutes of the deployment, as duplicate deposits began appearing in employee accounts. KiteBridge suspended automated payments a little more than an hour later, cutting off further duplicates. By then, however, the scope was already large. The duplicates affected dozens of corporate customers and thousands of individual reimbursements in a short window before automation was halted. KiteBridge’s early public communications understated what had happened. Its public incident page initially described the problem as delayed reimbursements, a far less alarming issue for finance officers watching payroll. The page did not acknowledge duplicate payments until six hours after the company had suspended automation. That delay is likely to draw scrutiny, particularly because KiteBridge had marketed reliability as a selling point. In April, its changelog for version 4.0 said the release had introduced “exactly-once payment protection” — industry shorthand for the guarantee that even if software retries a payment, the recipient is paid one time and only one time. Tuesday’s incident shows that protection failed when it was needed. Chief Executive Lena Ortiz has apologized for the error and said the company will publish a full technical report explaining how the fallback misfired and what it is changing to prevent a repeat. Recovery efforts are well underway. By Wednesday evening, KiteBridge said $2.3 million of the $2.8 million had been recovered, with the remainder still pending as the company works with employers and banks to reverse outstanding transfers. For affected businesses, the episode was disruptive even if temporary, forcing finance teams to reconcile books and explain unexpected deposits to staff. For KiteBridge, it is a test of trust in a business where precision is the product.

Anthropic: Claude Opus 5.5: # QueueZero's Proxy Assistant Turns Audience Remark Into CEO's To-Do List at Launch QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend meetings, produce notes and assign follow-up work to software agents. The launch demonstration also offered an unscripted look at how the system decides what counts as an instruction. During the demo, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy interpreted the remark as an instruction. It created two tasks assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. ## What Proxy Does Proxy is built to handle the administrative work that typically follows a meeting. It can join a call, generate notes and route follow-up items to software agents, rather than leaving attendees to track and distribute action items themselves. ## Built-In Safeguards According to QueueZero, Proxy requires confirmation before an external message is sent. Internal tasks, however, can be created automatically. Administrators can disable automatic task creation. That distinction is central to understanding what happened on stage. The tasks Proxy generated were internal, placing them in the category the system can create without first asking a user. In this case, the result was two items in the CEO's queue. Removing the unwanted one took a single deletion. Only one of the two tasks was ultimately marked complete. ## A Familiar Challenge for AI Assistants The episode highlights a broader problem for software that listens to conversations and acts on them: separating genuine requests from offhand comments, jokes and complaints. Everyday workplace speech is full of phrasing such as "someone should," which may signal a real action item or may simply be conversational. An assistant that acts on too much of what it hears risks filling task lists with noise. One that acts on too little may miss commitments that people expect to be tracked. QueueZero's design draws its firmest line at the boundary of the organization: actions that would reach outside parties require a human to approve them, while internal tasks can proceed on their own unless an administrator turns that capability off. For companies weighing the product, the practical question is whether automatic internal task creation saves more time than it costs in cleanup. The administrative setting gives organizations a way to choose a more conservative configuration. ## Pricing and Availability Proxy costs $30 per user per month. It enters public beta next week.

Meta: Muse Spark 1.3: # QueueZero Launches Proxy, an AI Assistant That Attends Meetings and Assigns Work QueueZero on Wednesday launched Proxy, an artificial intelligence meeting assistant designed to attend virtual meetings, produce notes and assign follow-up work to software agents. The announcement came at a company event in San Francisco where Chief Executive Evan Wu demonstrated the product live on stage. The system listens to meeting conversation, summarizes decisions and creates tasks it routes to integrated agents for execution. During the demonstration, a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy interpreted the remark as an instruction and created two tasks assigned to Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first, while the audience applauded. The moment underscored both the capability and the sensitivity of assistants that act on conversational cues without explicit commands. QueueZero said Proxy requires confirmation before an external message is sent. Internal tasks, however, can be created automatically based on what the assistant hears. Administrators can disable automatic task creation or require additional review, the company said. The incident highlights a broader challenge for makers of agentic software. As assistants gain permission to create work, schedule events and contact others, distinguishing casual remarks from intended instructions becomes critical. QueueZero framed its confirmation step for external communications as a safeguard against unintended actions, while leaving lower-risk internal task creation streamlined by default. Analysts say such controls will be closely watched as the category matures. Enterprises have expressed interest in tools that reduce manual note-taking and follow-up, but information technology leaders remain concerned about over-permissioned agents acting on ambiguous language in large meetings. Proxy costs $30 per user per month. It enters public beta next week, QueueZero said. The company plans to integrate the assistant with major videoconferencing platforms and project management tools at launch, with additional agent connections to follow. The company said early testers used Proxy to draft status updates and trigger support workflows. The company did not disclose customer numbers for the private preview. It said pricing includes transcription, summarization and task routing, with usage limits to be detailed at beta launch. Wu said the company believes meetings should produce outcomes without additional administrative work. QueueZero will publish documentation on configuration options and audit logs before the beta begins, he added.