Coral's AgentRadio lifted coding-agent accuracy to 62% at sixfold cost

Why it matters

AgentRadio shows that orchestration can extract a model-generation-sized gain from existing coding agents. The commercial tradeoff is a roughly sixfold increase in inference cost.

Caelum Forder and Peter Carroll, the founders of Coral AI Labs, have produced evidence that the architecture surrounding an AI model can matter as much as upgrading the model itself. Their researchers' AgentRadio paper, published July 30th, reports that four Claude Code agents running Anthropic's Opus 4.6 resolved 62.1% of a demanding codebase-understanding benchmark when they could exchange discoveries during execution. One Opus 4.6 agent resolved 32.3% in the researchers' runs, while a single agent using the newer Opus 4.8 scored 57.2% on Scale AI's public leaderboard.

The comparison is narrower than the headline result circulating after VentureBeat reported the research on August 7th. AgentRadio did not establish that Opus 4.6 is generally superior to Opus 4.8, or that four-agent systems consistently beat newer models. It beat one Claude Code and Opus 4.8 configuration on 124 code-comprehension tasks. Scale's current leaderboard already lists a single Claude Code agent using Opus 5 at 63.17%, slightly above AgentRadio's 62.1% result.

That quick reversal sharpens Coral's argument. Model releases can erase a benchmark lead in days. Coordination software can still generate large gains from models enterprises already use, especially on work that stretches across many files, tools and execution steps. The price is equally clear: AgentRadio's four-agent Opus 4.6 configuration averaged $19.45 per task, compared with $2.96 for one agent.

Forder is Coral's technical founder and CTO. Coral's current team page says he previously built AI infrastructure at IBM Watson, worked on AI safety at Conjecture and served as a founding engineer at multi-agent framework developer CAMEL-AI. Carroll, Coral's CEO, says he bootstrapped a game studio to 30 employees before funding Coral's early development and assembling the business around Forder's work on agent coordination. Lead researcher Xinxing Ren, a Brunel University London PhD graduate and former CAMEL-AI contributor, was one of seven authors on the paper alongside Qianbo Zang, Ziyan Wang, Forder, Suman Deb, Carroll and Zekun Guo.

What AgentRadio changes

AgentRadio addresses a mundane systems problem with large consequences: an agent executing a task usually cannot monitor a live conversation at the same time. Existing multi-agent designs often send several agents down separate paths and merge their answers later, or require every agent to stop at synchronized checkpoints before exchanging information. Both designs delay discoveries that could redirect work already underway.

Coral's system adds three communication operations to existing coding-agent harnesses: create_thread, send_message and wait_for_mention. An agent can open a named conversation, send a non-blocking message and run the mention watcher as a background operating-system process. Incoming messages appear between the agent's execution steps without stopping a command in progress.

The open-source implementation, released under Apache 2.0, does not require changes to Claude Code itself. Each agent communicates with a separate message server through thin shell scripts. The researchers ran one server and four Claude Code agents inside each benchmark container, using Modal and Harbor to manage the jobs.

Coral wrapped those primitives in a five-phase operating procedure. All four agents first inspect the repository independently. They then negotiate how to divide the question, execute their assigned investigations in parallel, review each other's evidence and approve a final answer assembled by one designated agent. During execution, an agent is instructed to post discoveries that affect another agent's assignment, contradict the agreed plan or expose a dead end.

The distinction between ordinary multi-agent delegation and AgentRadio sits in that execution phase. Under Coral's blocking configuration, an agent must spend a foreground step waiting for messages and therefore tends to work without listening. Under passive awareness, the watcher remains active in the background and supplies messages at the next step boundary.

Negotiation produced the largest gain

Coral's experiment separated the architecture into layers instead of comparing only the finished system with a single agent.

A lone Opus 4.6 agent resolved 40 of the 124 tasks, or 32.3%. Four agents with a basic division of labor resolved 49 tasks, lifting accuracy to 39.5%. Adding joint planning and cross-review raised the result to 64 tasks, or 51.6%. Moving message reception into the background produced the final 77-task result, or 62.1%.

The ladder matters because it shows that adding agents was the weakest intervention. Basic task division delivered 7.2 percentage points. Negotiation added 12.1 points, and passive awareness supplied another 10.5 points. The four-agent result came from a structured process that forced agents to negotiate assignments, expose evidence and revise work while it was still in progress.

The researchers repeated the architecture changes with DeepSeek V4 Pro inside the same Claude Code harness. The single-agent baseline scored 29%. Division raised it to 31.4%, negotiation reached 39.5%, and AgentRadio finished at 50.8%. The similar progression across two model families supports Coral's claim that the gain came from the coordination layer rather than an isolated quirk in Opus 4.6.

Coral also reported paired significance tests for the passive-awareness step. With Opus 4.6, background communication turned 15 previously failed tasks into passes while losing two tasks that the blocking system had passed. The corresponding counts for DeepSeek were 17 wins and three losses. The reported p-values were 0.0023 and 0.0026.

Those losses are important. Messages can distract an agent from a productive investigation, spread an incorrect conclusion or redirect attention toward a teammate's weaker theory. Agent communication introduces another source of context pollution even when the delivery mechanism works as designed.

The economics are harder than the benchmark

AgentRadio's strongest result cost about 6.6 times as much per task as the single Opus 4.6 run. Basic division averaged $5.38 per task, while adding negotiation raised the bill to $15.59. Passive awareness brought it to $19.45.

Coral tested whether the improvement simply came from spending more tokens by running the single agent six independent times and retaining the best answer. That approach cost $17.76 per task, close to the four-agent system's budget, and resolved 37.9% of tasks. AgentRadio reached 62.1% for $1.69 more. On the DeepSeek experiments, six independent runs cost $2.52 and scored 31.4%, while AgentRadio cost $2.46 and scored 50.8%.

The control supports a useful distinction between inference volume and organized inference. Repeating the same process produces additional attempts. AgentRadio allocates the budget across separate contexts, then uses negotiation and live messages to combine discoveries before the final answer is fixed.

Enterprise buyers still have to justify a jump from roughly $3 to nearly $20 for one codebase question. The calculation becomes plausible when an accurate answer prevents hours of engineering work, a production incident or a flawed migration. It becomes difficult for routine questions, high-volume automation and workflows where a person can cheaply verify the answer.

Anthropic reached a similar economic conclusion while describing its own multi-agent research system in 2025. Anthropic said multi-agent systems used about 15 times as many tokens as ordinary chat interactions and worked best on valuable, heavily parallelizable jobs. Anthropic also identified coding as a difficult case because software tasks contain dependencies and agents struggle to coordinate in real time. AgentRadio is a direct attempt to loosen that constraint rather than remove the underlying cost.

The benchmark rewards exhaustive understanding

SWE-Atlas Codebase QnA measures whether agents can understand unfamiliar production repositories before editing them. Its 124 tasks cover 11 repositories written in Go, Python, C and TypeScript. Questions span architecture, root-cause analysis, engineer onboarding, security and API behavior.

Static code search is insufficient. An agent receives a repository in a Docker container and may need to compile it, execute commands, reproduce behavior and trace a path across multiple files. Each answer is graded against an average of 12.3 factual criteria. A task counts as resolved only if every criterion passes, and modifying source files triggers an automatic failure.

The design makes SWE-Atlas a better test of long-running investigation than conventional coding evaluations centered on producing a patch. It remains one benchmark with its own grading choices. The AgentRadio experiments used Claude Opus 4.5 as the judge, fixed across all configurations, and ran each full 124-task configuration once. Coral separately repeated the four main Opus 4.6 configurations three times on a 30-task subset. AgentRadio averaged 64.4% there with a standard deviation of two percentage points, and its weakest run still exceeded the best run from the other configurations.

The public leaderboard also illustrates how quickly the reference point moves. The paper compared AgentRadio with Opus 4.8 because that was the strongest single-agent entry cited by the researchers. As of August 8th, Scale lists Opus 5 at 63.17%, alongside the Opus 4.8 result of 57.26%. Coral's result remains evidence for an architecture improvement. It is no longer the highest score among the entries being compared.

Live messages helped most when the plan was wrong

The largest gains appeared on architecture and system-design questions, where evidence tends to cross component boundaries. A naive four-agent division reduced Opus 4.6's resolved tasks in that category from 15 to 13. Negotiation raised the count to 24, and passive communication lifted it to 30. Splitting a tightly connected system into isolated assignments initially damaged performance; forcing the agents to reconcile those assignments reversed the loss.

Coral's MinIO case study shows the mechanism. A task required per-request server-side evidence, but none of the agents' initial plans mentioned enabling audit logging. In the blocking run, two agents separately encountered the logging issue and failed to turn it into shared evidence. One found the correct MINIO_AUDIT_WEBHOOK_ENABLE switch without proposing it to the others. The group later approved an incomplete conclusion.

In the passive run, an agent enabled the audit webhook and posted the resulting records while the other agents were still working. One piece of instrumentation became evidence available to the entire group, and the answer moved from passing 11 of 16 grading criteria to passing all 16.

A Grafana task exposed AgentRadio's boundary. Four grading criteria required the agents to establish that certain behavior did not occur. Neither configuration formed the correct negative conclusion, even after running relevant tests. Both passed five of nine criteria. A messaging layer can distribute an insight that one agent discovers. It cannot distribute an insight that none of them reaches.

Coral is turning research into an orchestration product

AgentRadio fits Forder and Carroll's broader bet on CoralOS, which Coral describes as an orchestration and governance layer for enterprise agent systems. CoralOS is built around connecting agents from different frameworks, monitoring their behavior and allowing existing APIs or MCP servers to participate in agent workflows. Coral's team page says the research has also been commercialized through a coding product called Coral Code.

That strategy places Coral below the model providers and above individual agent harnesses. An enterprise can swap models as benchmark rankings change while retaining the coordination, monitoring and approval logic that governs how agents work together. AgentRadio's design reinforces that position because its message server operates outside Claude Code and communicates through shell commands instead of modifying Anthropic's harness.

The benchmark result gives Coral a technically credible demonstration of that thesis. It also defines the burden facing orchestration vendors. They must prove that communication improves completed work after accounting for token spend, duplicated effort, propagated mistakes and the arrival of stronger base models.

AgentRadio cleared that bar on a deliberately difficult code-comprehension test. Its 29.8-point gain over one Opus 4.6 agent was much larger than the 5.6 points obtained by spending a similar budget on repeated independent runs. A newer single model has already caught the headline score. Coral's durable result is the controlled finding underneath it: agents working from the same model performed materially better when they could revise each other's investigations before execution ended.