Head to head: Z.ai: GLM 5.3 Prime vs MiniMax: MiniMax M3

Z.ai: GLM 5.3 Prime: # Arcwell launches Relay, an AI assistant for support agents Arcwell has launched Relay, a platform that uses artificial intelligence to examine customer support conversations and suggest answers to the agents handling them. The answers are grounded in what companies already have: Relay mines internal help documents and support transcripts for replies an agent can review and send. Answering is one of three jobs the platform does. Relay also identifies customers who may be about to cancel and creates summaries of conversations, so the next person to touch a ticket doesn't start from scratch. Chief Executive Priya Shah pitched the product as a fix for a frustration familiar to any support team. "Support agents spend too much time searching for information that their companies already have," she said. The information exists; the search is the cost. The performance case rests on a pilot that ran with 12 companies over six months. Agents using Relay closed tickets 14 percent faster, according to Arcwell — a gain that sounds small on a single ticket but adds up across everything a team closes in a day. The company did not release customer-retention results, even though spotting at-risk customers is one of Relay's three advertised capabilities. Relay costs $65 per agent per month. The speed comes with a data footprint. Relay requires access to support transcripts and internal help documents, which is where its answers and its churn flags come from. Arcwell says customer data is encrypted and is not used to train shared models, and companies can choose a 30-day or 90-day retention period. Those assurances are the ones any company will weigh before pointing an AI at its customer conversations. Relay is Arcwell's effort to help companies improve customer service with artificial intelligence. The ticket numbers from the pilot are the first public evidence that the approach works. The retention numbers, whenever Arcwell shares them, will be the more revealing test — of the product, and of the pitch.

MiniMax: MiniMax M3: # Arcwell Introduces Relay AI Platform for Support Agents Arcwell has announced Relay, an artificial intelligence platform that helps companies improve customer service by analyzing support conversations and suggesting answers to agents. The launch is a significant development for Arcwell and the broader customer-service industry, where artificial intelligence continues to change how support teams find information, assess customer risk and document interactions. Relay suggests answers to agents, identifies customers who may cancel and creates summaries of customer support conversations. The answer suggestions are designed to help agents draw on information their companies already have, while the summaries preserve the substance of an interaction. The cancellation feature brings potentially at-risk customers to agents’ attention. Together, these capabilities demonstrate how AI is changing the customer-service landscape. Arcwell tested Relay with 12 companies over six months. According to the company, agents using it closed tickets 14% faster. That result points to a productivity benefit, but the announcement did not include customer-retention data. It therefore remains unproven whether Relay’s use of AI and faster ticket closures affect how many customers leave. The platform costs $65 per agent each month. Priya Shah, Arcwell’s CEO, said, “Support agents spend too much time searching for information that their companies already have.” Relay requires access to support transcripts and internal help documents to suggest answers. That access gives it company-specific context, but it also raises questions about how customer information is handled. Arcwell says customer data is encrypted and is not used to train shared models. Customers can select a 30-day or 90-day retention period, giving them a choice over how long their data is kept. The launch marks an important new chapter for Arcwell as it seeks to help companies improve customer service with artificial intelligence. Its early case rests on faster ticket resolution, not demonstrated retention gains. That distinction, along with Relay’s data-access requirements and the privacy and retention terms Arcwell describes, will shape how companies evaluate the platform.

Z.ai: GLM 5.3 Prime: **C** The strongest lede is C because it is the only option that accurately reflects all the key facts: the product withdrawal, its price, the timeline, the number of affected accounts, and the specific nature of the exposure. Options A and D contain factual errors—A claims "private messages" were leaked and D claims "everything they had written" was visible to "hackers," but message contents were not exposed and no malicious access was found. Option B buries the actual news and uses promotional language ("innovative," "difficult decision") rather than reporting facts.

MiniMax: MiniMax M3: C

Z.ai: GLM 5.3 Prime: #

MiniMax: MiniMax M3: # Cloudnote Exposed User Prompts on Unauthenticated Debug Page *The AI note-taking service says excerpts from 1,842 sessions were visible for roughly four hours before a researcher flagged the issue.* Cloudnote, an artificial-intelligence note-taking service with 3.2 million registered users, left an internal debugging page accessible without authentication on Monday, briefly exposing short excerpts of customer prompts. The page, designed for internal troubleshooting, was reachable for roughly four hours, according to the company. Server logs indicate it contained fragments from 1,842 user sessions, with each excerpt including up to 200 characters of a submitted prompt. Cloudnote removed the page 19 minutes after an outside researcher reported the problem. Even partial prompts can carry meaningful context — proprietary business questions, personal medical details, code snippets or fragments of confidential work — making the exposure notable despite its brevity. The episode highlights the persistent risks of leaving debugging tools connected to live production systems, a recurring source of data incidents across the technology industry. The company said it has found no evidence that anyone other than the researcher viewed the page, though it cautioned that its investigation is not yet complete. Cloudnote also noted that full documents, passwords and payment information are stored on separate systems and were not present on the exposed page.

Z.ai: GLM 5.3 Prime: # KiteBridge Bug Sent $2.8 Million in Duplicate Payments to 91 Corporate Customers A faulty software update deployed by expense-management firm KiteBridge generated 13,400 duplicate reimbursements totaling $2.8 million on Tuesday, affecting 91 corporate customers and raising questions about a payment safeguard the company had promoted as protection against precisely this kind of failure. KiteBridge's software is used by companies to review and pay employees' expense claims — travel, meals, supplies and the like. Version 4.2 of the platform went live at 8:45 a.m. Tuesday, and by the time the company suspended automated payments just over an hour later, the duplicates had already gone out. The company had recovered $2.3 million as of Wednesday evening, leaving about $500,000 pending. Employees lost nothing — no money was withdrawn from their accounts — because the duplicates were additional payments funded by employers. The cash-flow burden, until the funds were clawed back, fell on the affected companies. ## How a payment timeout became a $2.8 million problem The failure began with an everyday hazard of payment systems: not knowing whether a transaction succeeded. When KiteBridge approves a reimbursement, it sends an instruction to a payment processor, the intermediary that moves money into an employee's bank account. Sometimes that exchange stalls — a "timeout" — and the software never learns whether the payment actually went through. The usual remedy is to try again. But the bug in version 4.2 sat in KiteBridge's fallback system, the backup code that takes over when the primary path fails. It treated each retry as an entirely new reimbursement request rather than a repeat of the original one, so payments whose status was unknown were issued again. Across 91 customers, that produced 13,400 duplicate reimbursements worth $2.8 million. The irony is sharp. A changelog entry from April, when KiteBridge released version 4.0, said the software had gained "exactly-once payment protection" — industry shorthand for a guarantee that every payment executes a single time, even amid errors and retries. Whether version 4.2 broke that safeguard, bypassed it or exposed a flaw in it is not clear; the company has not said. ## A fast shutdown, a slow disclosure The first customer complaint arrived at 9:12 a.m., less than half an hour after the rollout began. At 10:03 a.m., KiteBridge suspended automated payments, roughly 80 minutes after version 4.2 went live, halting further duplicates — though not before thousands had already been paid. The company's public account lagged. Its incident page initially described the problem as delayed reimbursements, an understated description of a failure in which extra payments had already been made. Only about six hours after automated payments were suspended — late Tuesday afternoon — did the page disclose that duplicate payments had occurred. ## Open questions CEO Lena Ortiz has apologized and said KiteBridge will publish a technical report on the failure. That report will have to answer the central question of the episode: how a system advertised as offering exactly-once protection came to pay thousands of expenses more than once. For now, recovery continues. With $2.3 million returned and about $500,000 outstanding, some of the 91 affected companies are still waiting for money they never meant to send — and for a fuller accounting of the hours between KiteBridge's internal response and its public one.

MiniMax: MiniMax M3: # KiteBridge Bug Triggers $2.8 Million in Duplicate Expense Reimbursements A routine software update at expense-management firm KiteBridge has resulted in roughly $2.8 million in duplicate reimbursements flowing to employees at 91 corporate customers, despite the company having publicly advertised an "exactly-once payment protection" feature less than two months earlier. The episode has prompted public criticism over the firm's delayed and initially incomplete disclosure of what went wrong. ## How It Unfolded At 8:45 a.m. on Tuesday, KiteBridge pushed version 4.2 of its platform to production. Within half an hour, the first customer complaint landed in the company's support queue at 9:12 a.m. Behind that ticket was a fault in the software's payment-orchestration layer: when the third-party payment processor timed out on a transaction, KiteBridge's fallback logic treated the original request as failed and the processor's automatic retry as a brand-new reimbursement. In normal conditions, that fallback exists to ensure payments are not lost; on Tuesday it did the opposite. The misfiring system ultimately generated 13,400 duplicate reimbursements before engineers suspended automated payouts at 10:03 a.m., roughly an hour and 18 minutes after the first complaint. Affected disbursements totaled $2.8 million, distributed across 91 corporate clients of KiteBridge's expense platform. By Wednesday evening, the company said it had clawed back $2.3 million of the overpayments. The remainder remained pending as of the latest update, with KiteBridge working with its banking partners to reverse the straggling transfers. ## A Promising Changelog Meets a Costly Bug The incident is particularly awkward for KiteBridge because of what the company said just weeks ago. An April changelog announcing version 4.0 described the release as introducing "exactly-once payment protection," language that, in distributed-systems parlance, suggests a transaction will never be applied more than once no matter how many retries or failures occur along the way. Tuesday's failure suggests that guarantee was not actually held under realistic failure conditions, such as a third-party timeout. ## Communication Questions KiteBridge's public handling of the disclosure has also drawn scrutiny. The company's incident page first described the issue as "delayed reimbursements," language that framed the problem as a slowdown rather than a duplication error. It was not until roughly six hours after automated payments were suspended — late Tuesday afternoon — that the page acknowledged duplicate reimbursements had been issued and quantified the exposure. In a statement late Wednesday, CEO Lena Ortiz apologized for the incident and said the company would publish a full technical report explaining the root cause, the gap in the version 4.0 guarantees, and the steps being taken to prevent recurrence. ## Who Actually Paid One point of relief for the affected workers: KiteBridge stressed that no employee had money pulled directly from a personal account. The duplicate payments were funded by the 91 corporate customers, which had pre-funded reimbursement pools with KiteBridge. Those employers are now waiting for reversals or credits to restore their balances. For KiteBridge, the immediate work is financial: recovering the remaining $500,000 and rebuilding trust with customers whose expense accounts briefly paid out twice. The longer-term work — restoring credibility around the "exactly-once" promise — will depend heavily on what Tuesday's forthcoming technical report actually says.

Z.ai: GLM 5.3 Prime: # Foundry Hall to Close Oct. 31 After Losing Its Largest Sponsor *The nonprofit accelerator helped launch 212 startups over six years. It could not survive the loss of one sponsor.* MILWAUKEE — Foundry Hall, the nonprofit startup accelerator that has anchored the city's entrepreneurial scene since 2018, will close its doors on October 31. For six years, it offered early-stage companies the things a city can't easily improvise: shared workspace, a structured program, and an address where a founder could be taken seriously. The cause of the closing is money. Blank River Bank, Foundry Hall's largest corporate sponsor, supplied 38 percent of its annual budget and this year declined to renew. A gap of that size is not the kind a small nonprofit quietly absorbs. A spokesperson for the bank said the decision reflects a change in its community investment priorities and is "unrelated to Foundry Hall's performance." The reassurance is precise. It is also beside the point for an organization whose performance — 212 startups supported over six years, companies that now employ 1,140 people — was the whole argument for its existence. Executive director Camille Ross said Foundry Hall has enough money to operate through October but not enough to run another full program. Behind the closing is a long, fruitless search for a sponsor to replace the bank. "We kept waiting for the replacement check that never came," Ross said. The staff will keep the doors open through Halloween, running out the calendar on money already in hand. The closing will cost six employees their jobs. It will also upend the twelve startups currently using Foundry Hall's workspace, companies that must find somewhere else to go while trying to keep building. For a young company, an unexpected move is more than logistics. Momentum is hard to relocate. Dev Malik, a founder who got his start at the accelerator, remembers what the building replaced. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." Alumni like Malik can measure that distance in years. The founders still inside the building have weeks. Deputy Mayor Susan Cole said the city is speaking with building owners and local universities about space for the current companies — for now, conversations rather than commitments. No arrangement has been announced, and the deadline is fixed. By one measure, Foundry Hall's legacy is easy to state: 212 startups, 1,140 jobs. By another, it resists arithmetic — first investor meetings, desks that replaced back tables, six people now looking for work. On October 31, the building closes. What it started doesn't.

MiniMax: MiniMax M3: When Dev Malik walks through Foundry Hall one last time, the lobby still smells faintly of fresh coffee and late-night whiteboard sessions. "My first investor meeting happened in that building," Malik said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." Foundry Hall, a nonprofit startup accelerator in Milwaukee, will close its doors on October 31. When it opened in 2018, the organization set out to give early-stage founders somewhere to work that they otherwise could not afford. Over seven years, it supported 212 startups that now collectively employ more than 1,140 people across the region. The closure hinges on a single signature that did not arrive. Blank River Bank, the accelerator's largest corporate sponsor, had supplied 38 percent of its annual operating budget. This year, the bank declined to renew. A spokesperson for Blank River Bank framed the decision as a realignment of community investment priorities, emphasizing that it was unrelated to Foundry Hall's performance — language that did little to soften the impact for the six employees who will lose their jobs at the end of October, or for the twelve startups currently sharing workspace inside the building. "We kept waiting for the replacement check that never came," executive director Camille Ross said. Ross has led the accelerator since its founding, watching it grow from a single shared office into a regional hub. The organization has enough funding to operate through October, but not enough to launch another full cohort. When the building empties, the companies inside will need to find new offices, and the steady pipeline of new ventures the accelerator once fed into the local economy will go quiet. Deputy mayor Susan Cole said the city is already speaking with private building owners and local universities about temporary space for the displaced founders — an arrangement that, if it materializes, would buy them time but not permanence. For Malik, the loss runs deeper than the lease itself. He met his first business partner in a third-floor conference room, hired two of his current employees through introductions made in the shared kitchen, and credits the accelerator's mentorship program with helping him survive a pandemic-era cash crunch. That room, like the back table he once pitched from, will soon be empty. The founders who built their companies inside Foundry Hall's walls are now searching for what comes next.

Z.ai: GLM 5.3 Prime: 5

MiniMax: MiniMax M3: 6

Z.ai: GLM 5.3 Prime: # 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.

MiniMax: MiniMax M3: # 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.

Z.ai: GLM 5.3 Prime: # The Priciest Model Didn't Win: A 200-Ticket AI Support Benchmark When five language models were tested head-to-head on 200 anonymized customer-support tickets, the most expensive one failed to justify its price. Model North, at an estimated $1.90 per 1,000 tickets, averaged 84.2 points. Model Vale averaged 83.8 — a difference small enough that researchers called it a statistical tie — while costing $0.74 per 1,000 tickets, roughly 39 percent of North's price, and returning answers 18 percent faster. That's the benchmark's clearest finding: near the top of the quality range, price stopped predicting performance. A 2.6x premium bought a tie. The full results, for reference: | Model | Average score | Est. cost / 1,000 tickets | |---|---|---| | North | 84.2 | $1.90 | | Vale | 83.8 | $0.74 | | Ember | 79.1 | $0.31 | | Kite | 76.4 | $0.22 | | Harbor | 71.0 | $0.08 | ## How the test worked The setup was simple. Researchers took 200 real tickets from a single software company, anonymized them, and had each of the five models draft one proposed answer per ticket. Three experienced support managers then graded every answer blind — none of them knew which model had produced what — scoring each response on a blend of factual correctness, usefulness, and tone. Two design choices are worth pausing on. One answer per ticket means the scores measure first-draft quality; a production pipeline with retries, editing, and human review could produce different results. And blind grading is what makes the North–Vale tie credible: whatever the managers think of particular model brands, those impressions couldn't leak into the scores. ## A tie with a fork in it The overall tie hides a split that matters more to buyers than the average does. North performed best on billing disputes. Vale performed best on technical troubleshooting. A support team whose queue leans heavily toward one of those has a practical winner even though the headline scores tie; a mixed queue pushes the decision to cost and speed. Arguably the benchmark's most useful output isn't the ranking but that category detail. The lesson for buyers is to match the model to your ticket mix, not to the leaderboard. ## What the bargain tier gives up Below the top two, quality falls off faster than price does. Ember averaged 79.1 at an estimated $0.31 per 1,000 tickets, Kite 76.4 at $0.22, and Harbor 71.0 at $0.08. The only significance result reported was for North versus Vale, so the ordering among the cheaper models shouldn't be treated as settled; on 200 tickets, a few points can be noise. But Harbor sits roughly 13 points behind the leaders — a gap big enough to matter for customer-facing work no matter how it's priced, and a discount never makes a wrong answer cheap. That said, the absolute stakes are smaller than the ratios suggest. Even at North's rate, a million tickets a year would cost about $1,900 in generation fees — pocket change next to agent salaries. For high-volume teams, the real question is whether a cheap model's quality floor clears the bar, not whether its savings compound. ## Where every model failed One result held across all five: they performed poorly whenever a ticket's answer depended on an image attachment. Support queues are full of screenshots — error messages, broken interfaces, damaged products — so this is a shared blind spot rather than a differentiator. Whichever model you choose, image-dependent tickets remain a natural lane for human agents or a dedicated multimodal workflow. ## Read the fine print - **Scope.** 200 tickets, one English-language software company. Scores may not transfer to other industries, languages, or ticket styles. - **Costs are estimates.** Real invoices depend on prompt length, context size, and pricing changes. - **One draft, three graders.** The study doesn't measure consistency — how often each model produces a clunker — or report how closely the graders agreed. - **Drafting only.** Models wrote answers but took no actions: no tool use, no agentic steps like issuing a refund, no multilingual support tested. Much real support work is doing, not just writing. ## The takeaway If you're picking a model for support drafting, both default instincts — buy the most expensive, or buy the highest scorer — need checking. Here the top two were indistinguishable on quality, and the cheaper one was faster. The right move is to run your own bake-off on your own ticket mix, weighted toward the categories you actually receive, and to route image-dependent tickets elsewhere. In this benchmark, first place was a tie; the interesting decision happened somewhere below the average.

MiniMax: MiniMax M3: # Vale Matches the Quality Leader—and Wins on Efficiency The clearest result is not that Model North won, but that Model Vale essentially matched it while answering faster and costing less. North posted the highest average composite score, 84.2, while Vale scored 83.8. The 0.4-point gap was not statistically significant, so the test provides no reliable evidence that North produced better answers. Vale answered 18% faster, and its estimated cost was $0.74 per 1,000 tickets versus North’s $1.90. For a support organization, that combination makes Vale the strongest overall candidate: not a proven quality winner, but a model that preserved the leader’s measured performance while improving efficiency. ## What Was Tested Researchers gave five models the same 200 anonymized customer-support tickets. Each produced one proposed reply per ticket, and three experienced support managers graded the replies blindly, meaning they did not know which model generated each answer. Blind grading helps reduce brand and expectation bias. The score combined factual correctness, usefulness, and tone, making it a practical measure of draft quality rather than a pure test of factual accuracy. That design supports a fair comparison, but it also has boundaries. A composite score can conceal tradeoffs: a model might lose points on accuracy while gaining on tone, and the average would not reveal it. The nonsignificant North–Vale gap means the study did not demonstrate North’s superiority; it does not prove that the models are exactly equivalent everywhere. ## Quality Has a Price North and Vale form the apparent quality tier. The other three models were cheaper but scored lower, showing a broad cost–quality tradeoff. Kite and Ember occupied the middle, with average scores of 76.4 and 79.1 and estimated costs of $0.22 and $0.31 per 1,000 tickets. Harbor reached the lowest cost, $0.08, but its average score fell to 71.0. The study does not establish the exact shape of that tradeoff or whether every difference among the lower-ranked models was statistically significant. Still, the lowest-cost option did not come without a visible quality penalty. Category results add useful nuance. North performed best on billing disputes, where precise policy explanations may matter most. Vale performed best on technical troubleshooting, suggesting strength in organizing diagnostic steps. A support team could therefore test targeted routing—billing cases to North, troubleshooting cases to Vale—if the pattern holds on a larger sample. ## Image Attachments Remain a Shared Weakness More important, all five models performed poorly when a ticket depended on an image attachment. Screenshots can contain the decisive error message, account state, or device indicator, so a strong general score should not be assumed to cover visual support. Production systems need a fallback: request text, escalate to a human or multimodal workflow, and measure those handoffs separately. ## Limits on the Conclusion The scope is narrow. All tickets were in English and came from one software company, so the results cannot automatically be generalized to other products, industries, customer populations, or languages. The sample may also underrepresent ticket types outside that dataset. A larger, stratified evaluation would be needed to confirm the category patterns and the prevalence of image-dependent cases. Researchers also tested only proposed answers. They did not evaluate agentic actions, tool use, or multilingual support, leaving open whether a model can retrieve an order, update a CRM record, execute a troubleshooting step, or assist across languages. Those are consequential gaps because live support depends on actions and integrations, not just wording. One answer per ticket also says little about consistency across repeated attempts. Finally, combined scoring limits diagnosis of the capability behind each result, and estimated costs will change with traffic, retries, human review, and escalation. ## Practical Verdict For now, Vale is the most defensible default. It matched North’s quality within the limits of the evidence, delivered answers 18% faster, and cut estimated cost by roughly 61%. North remains compelling for billing-heavy workloads, and Vale appears especially promising for technical cases, but a 0.4-point difference should not decide the rollout. The broader lesson is that support automation involves quality, latency, and operating cost together. Before deployment, teams should rerun the evaluation on their own tickets, score the dimensions separately, include image-dependent cases, and track corrections and escalations. This benchmark identifies two viable leaders; it does not yet justify an autonomous support agent.

Z.ai: GLM 5.3 Prime: *

MiniMax: MiniMax M3: # Startup Spotlight: LedgerLark The mismatch started with a single copied digit. A six-person operations team at a regional brokerage spent three days reconciling a trade that had gone sideways because an account number had been duplicated somewhere between the order ticket and the custodian's records. Mina Batra, then seven years into her own operations role at a similar firm, kept thinking about that story long after it was resolved. That experience is now the founding insight behind LedgerLark, an early-stage company building AI agents for post-trade operations at small investment firms. The agents compare trade confirmations, custodian records, and internal ledgers, then propose corrections for human review. They do not execute trades, move money, or touch funds in any way. Batra teamed up with Owen Carr, who spent years building reconciliation systems at a large bank. The pairing reflects a common pattern in financial infrastructure: someone who lived with the operational pain on the customer side, paired with someone who understands the technical plumbing. ## What the product does Post-trade operations — the work that happens after a trade is executed — is a famously unglamorous corner of finance. It involves matching records across multiple systems, flagging discrepancies, and chasing down why one number doesn't agree with another. At smaller firms, this work often falls to small teams wearing many hats. LedgerLark's pitch is straightforward: automate the matching and exception-handling work that humans currently do manually, then route suggested fixes to a human for approval. The company has been deliberate about positioning the product as assistive rather than autonomous, partly to address the compliance sensitivities that come with anything touching financial data. ## Early traction, with caveats Five firms are paying customers today, with annual contracts ranging from roughly $36,000 to $110,000. The company points to one customer that reduced its unresolved daily exceptions from about 70 to 18 after deploying the software, though independent verification of that figure isn't available, and the result likely reflects a mix of automation and changed internal workflows. It's a reasonable starting point for a company that began shipping product relatively recently, but five customers is a small base, and contract sizes vary widely — a sign that pricing or scope is still being calibrated. ## Competition and the road ahead LedgerLark isn't the only group chasing this problem. Established reconciliation vendors have decades of installed base, and large banks have built internal systems like the ones Carr once worked on. The opportunity for LedgerLark is in the long tail of small and mid-sized firms that can't afford enterprise-tier tools or dedicated engineering teams. Even with a product in market, the company faces a familiar obstacle in financial services: lengthy security and procurement reviews at regulated institutions can stretch sales cycles and complicate forecasting. The company has raised a $3.5 million seed round led by Motive Row, which should give it runway to navigate those cycles, though it's early to say whether the funding level matches the ambition of selling into an industry known for slow adoption. ## Outlook What's notable about LedgerLark is how narrow and concrete the problem is. There's no pitch here about transforming finance or replacing back-office staff. The story is closer to: a tedious class of errors is expensive, and software can help catch them faster. Whether that translates into a durable business depends on factors the founders can't fully control — sales cycles, customer churn in a still-tiny base, and the willingness of incumbents to leave the middle market alone. For now, LedgerLark is a small company with a specific bet and a few reference customers. Both the opportunity and the uncertainty around it are worth taking seriously.

Z.ai: GLM 5.3 Prime: # Raincheck raises $8 million to help cities get ahead of floods *The Austin company combines drainage maps with rainfall forecasts to show local governments where to send crews before a storm arrives.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised an $8 million seed round, the company announced today. 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. Raincheck details its forecasting approach in [a methodology overview](https://example.com/raincheck-methodology). Basin Ventures led the round, with existing investor Long Mile participating. The funding brings Raincheck's total raised to $10.5 million. CEO Ana Flores said: > "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. The company'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 the new funding will help it hire eight engineers and expand into Florida.

MiniMax: MiniMax M3: # Raincheck Raises $8 Million to Expand Flood-Risk Software for Cities *Austin-based startup maps drainage systems and forecasts where crews should respond before a storm hits.* Raincheck, an Austin-based company that builds flood-risk software for local governments, has raised $8 million in a seed round announced today. The company maps municipal drainage systems and combines those maps with rainfall forecasts, according to [its published methodology](https://example.com/raincheck-methodology). Cities use the software to decide where to send maintenance crews ahead of a storm. The round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has raised $10.5 million to date. > "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." *— Ana Flores, CEO of Raincheck* Raincheck currently works with seven cities in Texas and Louisiana, though the company declined to name 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 support hiring eight engineers and an expansion into Florida.

Z.ai: GLM 5.3 Prime: # QueueZero Launches Proxy, an AI Meeting Assistant That Assigned Its CEO a To-Do List *The new assistant treats what it hears as instructions — as its own chief executive learned mid-demo.* QueueZero launched an AI meeting assistant called Proxy on Wednesday, a tool designed to attend meetings, produce notes and assign follow-up work to software agents that carry out tasks on their own. The launch demonstration produced an unplanned moment that illustrated both the product's appeal and its risks. During the demonstration, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy took the remark at face value. Interpreting it as an instruction, the assistant created two tasks and assigned both to Evan Wu, QueueZero's chief executive: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. The record does not show whether the talking stopped. ## What Proxy does Proxy is the latest entry in a fast-growing category of AI tools that aim to move meeting software beyond transcription. The assistant can attend a session, generate notes and then convert the discussion into follow-up items. Its distinguishing feature is what happens next: rather than leaving action items in a document for humans to distribute, Proxy hands them to software agents that carry them out. Wednesday's incident also clarified the product's guardrails. QueueZero says Proxy requires confirmation before an external message is sent, which means the assistant cannot contact people outside an organization without a human approving the message. Internal tasks are different. Those can be created automatically, with no confirmation step — which is how a customer's aside became two items on a chief executive's to-do list. Administrators who want tighter control can disable automatic task creation altogether. ## A familiar trade-off The episode is a compact illustration of the central tension in agentic AI, in which systems do not merely answer questions but take actions on a user's behalf. Autonomy is the value proposition; literal-mindedness is the known failure mode. A tool that treats an overheard remark as a directive is, by one reading, working exactly as designed — and organizations adopting such systems will have to decide whether that is a feature or a liability. ## Pricing and availability Proxy costs $30 per user per month and enters public beta next week. By then, administrators will be able to decide for themselves how much autonomy to give an assistant that has already put it to use.

MiniMax: MiniMax M3: # QueueZero's AI Meeting Assistant Creates Tasks Mid-Demo After Audience Quip QueueZero unveiled an artificial intelligence meeting assistant called Proxy on Wednesday, pitching the software as a way to offload the administrative work that typically follows a calendar invite. The tool can join a meeting, generate notes, and assign follow-up tasks to software agents connected to a team's project management system. The launch demonstration took an unexpected turn when a customer in the audience shouted, "Someone should tell Evan to stop talking and send us the pricing sheet," referring to QueueZero chief executive Evan Wu. Proxy, listening for actionable items, parsed the remark as a directive and promptly created two tasks in the company's task board. Both were assigned to Wu: "Stop talking" and "Send pricing sheet." Wu, presenting at the time, laughed and addressed the room. He said he would complete the second task and skip the first. He then uploaded the pricing document and deleted the unwanted assignment. The moment, captured by several attendees, was shared widely on social media within hours of the event. The episode highlighted a question that has followed the broader category of AI meeting assistants: how much initiative should such software be given? QueueZero says Proxy requires explicit human confirmation before any external message leaves the platform, including emails or chat pings to people outside the meeting. Internal tasks, however, can be generated automatically when the assistant identifies what it interprets as a clear instruction. Administrators who want tighter guardrails can disable automatic task creation entirely, the company said, requiring a person to approve each entry before it lands on a shared board. Wu described the layered controls as a deliberate choice, designed to make the tool useful without making it reckless. Proxy will enter public beta next week. The company priced the product at $30 per user per month, with volume discounts available for teams above fifty seats. QueueZero said early customers include software firms, marketing agencies, and a handful of professional services companies. The market for meeting software has grown crowded since the introduction of large language models, with established players and startups racing to bundle transcription, summarization, and workflow automation into a single product. Analysts at Gartner have forecast that spending on AI-enabled collaboration tools will reach several billion dollars annually within three years, though they caution that reliability and oversight remain open issues. For QueueZero, Wednesday's demonstration may have been the most effective advertisement the company could have asked for, even if the moment was unplanned.

Z.ai: GLM 5.3 Prime: #

MiniMax: MiniMax M3: # HelioDock launches Night Shift for remote warehouse robot monitoring *HelioDock’s fleet-management software is set for general availability Sept. 15 after a 14-month private beta, as the autonomous-robotics company raises $24 million.* HelioDock launched Night Shift on Thursday, software that lets warehouse operators monitor fleets of autonomous robots remotely, as the warehouse-robotics maker announced a $24 million Series B led by North Coast Ventures. The product extends the company’s work from making autonomous machines into software for managing them in warehouses. Night Shift is intended to reduce the need for managers to reach a facility when a robot encounters trouble. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” said HelioDock CEO Mara Velez. Her example captures the operational problem behind the software: an issue with one machine can require attention at a site, even when managers are not on the premises. Night Shift enters a field with existing alternatives. Relay Robotics and YardOS already offer competing fleet-management products, giving operators a choice among several vendors. HelioDock’s launch adds another option for supervising robot fleets remotely rather than introducing remote fleet management as a category limited to one supplier. The financing announced alongside the launch brings HelioDock’s total funding to $39 million. North Coast Ventures led the $24 million round. The two announcements place hardware and software at the center of the company’s current expansion, with the new product adding remote oversight to a business built around autonomous warehouse robots. Before Thursday’s launch, Night Shift underwent a 14-month private beta involving nine logistics companies and a total of 310 robots. The duration and scale of the test define the scope of the evidence behind the release. Across the beta group, HelioDock said, customers experienced 17% fewer unplanned robot stoppages. The 17% figure is a company-supplied claim and has not been independently verified. It also measures a specific outcome: unplanned robot stoppages. That is narrower than a measure of total downtime, fleet productivity or warehouse output. As an aggregate result from the beta, the figure does not by itself establish the performance of Night Shift at every participating facility, though it is the principal performance claim associated with the test. Feedback from one customer shows why the software’s operating conditions matter. Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. His comments point to a practical dividing line: Night Shift’s remote visibility depends on the network connection available at a given facility. When connectivity fails, HelioDock robots return to the behavior configured locally at the facility, and remote control becomes unavailable. The local configuration remains the basis for robot behavior during the outage, but managers lose the ability to control a machine from elsewhere. Night Shift therefore supplements local operation with remote visibility when a connection is available; it does not provide continuous remote control through a network failure. Night Shift is scheduled to become generally available Sept. 15. Under its announced pricing, operators will pay $90 per robot per month, a model that ties the software’s monthly cost directly to fleet size. The recurring fee also gives HelioDock a software offering alongside its robot business as Night Shift moves beyond the nine-company private beta.