--- title: "CooperBench and the curse of coordination: why coding agents do worse together · Cork" description: "CooperBench (arXiv:2601.13295) found coding agents do worse working together than one agent doing both tasks. The paper's numbers, stated precisely, and where Cork stands: a CooperBench run is planned." canonical: "https://corkcli.com/blog/cooperbench-curse-of-coordination" last_updated: "2026-11-14" published: "2026-10-19" author: "Kevin Vera" --- # CooperBench and the curse of coordination: why coding agents do worse together CooperBench (arXiv:2601.13295) found coding agents do worse working together than one agent doing both tasks. The paper's numbers, stated precisely, and where Cork stands: a CooperBench run is planned. By Kevin Vera · October 19, 2026 · [Research](https://corkcli.com/blog/c/research) · 7 min read In January 2026, researchers from Stanford and SAP published CooperBench, a benchmark that asks a blunt question: can two coding agents work together as well as one agent doing both jobs? On average, no. The paper calls it "the curse of coordination". This post covers what CooperBench measures, what it found, how its headline number is stated in the paper, why it matters to anyone running several agents on one repository, and where Cork stands: we have not run CooperBench yet, and a run is planned. ## The paper "CooperBench: Why Coding Agents Cannot be Your Teammates Yet" by Arpandeep Khatua, Hao Zhu, Peter Tran, Arya Prabhudesai, Frederic Sadrieh, Johann K. Lieberwirth, Xinkai Yu, Yicheng Fu, Michael J. Ryan, Jiaxin Pei and Diyi Yang. arXiv:2601.13295, submitted January 19, 2026 and revised January 26, 2026. "CooperBench comprises 652 tasks constructed from 12 popular open-source libraries across Python, TypeScript, Go, and Rust." In each task, two features are implemented against the same repository, and the paper defines two ways to do it: "each of the two agents are assigned a feature to implement, which will be called the Coop setting to distinguish from the Solo baseline. In the Solo baseline, the two tasks are assigned to one agent." In the Coop setting, the agents work separately and their only channel is a messaging tool; afterwards, their patches are merged with `git merge-file` and the result is judged by the features' tests. A task succeeds only if the work merges and both features are correct. The paper evaluated agents built on GPT-5, Claude Sonnet 4.5, MiniMax-M2 and two Qwen models. ## The curse of coordination, precisely The paper states the gap in more than one way, and they are worth keeping apart: - **Abstract:** "agents achieve on average 30% lower success rates when working together compared to performing both tasks individually, across the full spectrum of task difficulties." - **Introduction:** "GPT-5 and Claude Sonnet 4.5 based agents achieve only 25% with two-agent cooperation on CooperBench, which is around 50% lower than a 'Solo' baseline which uses one agent to implement both features." - **Section 4:** "This coordination gap is as large as 50% in the leading models: GPT-5, Claude Sonnet 4.5, and Minimax M2." - **Appendix C:** "On average, 41% of Solo capability is lost when agents must coordinate (pooled retention 0.59)." The appendix's counts of tasks solved: | Model | Solo | Coop | Share of Solo kept | | --- | --- | --- | --- | | GPT-5 | 315 | 183 | 0.64 | | Claude Sonnet 4.5 | 307 | 168 | 0.60 | | MiniMax-M2 | 236 | 91 | 0.46 | | Qwen3-Coder | 141 | 87 | 0.63 | | Qwen3 Instruct | 41 | 30 | 0.68 | | Pooled | 1,039 | 558 | 0.59 | So the commonly quoted "about 30% lower" is the abstract's average; the paper's own appendix puts the pooled loss at 41%, and the strongest models lose up to about half. Whichever figure you use, the direction is the same: two agents sharing the work did worse than one agent doing all of it. It gets worse with more agents. On a 46-task subset, "performance drops from 68.6% with 2 agents to 46.5% with 3 agents and further to 30.0% with 4 agents." ## Why talking did not fix it The obvious remedy is communication, and the paper tested it: "The difference between 'with comm' and 'no comm' settings is not statistically significant." Messaging did reduce merge conflicts, and it was not cheap: "agents spent as much as 20% of the steps in communication." The abstract names three failures: "(1) communication channels become jammed with vague, ill-timed, and inaccurate messages; (2) even with effective communication, agents deviate from their commitments; and (3) agents often hold incorrect expectations about others' plans, observations, and communication." Reviewing 50 failed traces, the authors attribute 42% to expectation (an agent was told what its partner was doing and acted as if it were not), 32% to commitment and 26% to communication. Their diagnosis is the part we find most useful: "These failures suggest current models lack reliable representations for (i) partner state (what the other agent has actually changed), (ii) checkable commitments (contracts verifiable after merge), and (iii) cross-branch integration reasoning." ## Why this is Cork's problem Cork exists because running several coding agents on one codebase goes wrong in exactly these ways. Its design bets on the first item in that diagnosis: instead of hoping an agent remembers what its partner said, it hands the agent its partner's actual state, from the code. - **Partner state from the code, not from chat.** Each agent is briefed on the exact signatures, callers and recent changes of the code it is about to touch, and hears on its next step when another agent renames or re-signatures something it depends on. - **Edits held against live work.** The [write gate](https://corkcli.com/blog/write-gate) holds an edit against code another agent is changing right now, until the agent has the new version. - **Integration checked at landing.** `git push` rebases onto teammates' commits, reruns the tests and pushes in one step. - **Messages tied to code.** Agent messages can be anchored to a symbol, and the recipient is told whether that code has changed since the message was sent. Whether this closes CooperBench's gap is an open question. We have not run it. ## What our own studies show, and what they do not Cork's published studies measure something related but different: the same tasks run by groups of agents with Cork and without it, counting tokens and checking correctness with each task's own tests. For example: - [Study 7](https://corkcli.com/benchmarks/study-7-scaling-to-sixteen-agents): 1 to 16 Claude Code agents at once on one repository, five rounds at each size. Agents used 24% fewer tokens with Cork at one agent and 47% fewer at sixteen; 151/155 tasks were correct with Cork and 151/155 without; at sixteen agents, rejected pushes averaged 0 with Cork vs 40.6 without, and merge conflicts 0.4 vs 7.6. - [Study 5](https://corkcli.com/benchmarks/study-5-twelve-agents-53-percent): twelve agents on hard tasks used 53% fewer tokens with Cork's paid plan than without. - [Study 6](https://corkcli.com/benchmarks/study-6-four-people-twelve-agents): four people, three agents each, four rounds: 29% fewer tokens, 47/48 tasks correct each way. These are not CooperBench results. Our tasks, repositories and harness are our own, correctness was about the same with and without Cork, and none of our studies compares agents working together against one agent working alone, which is CooperBench's question. All studies, methods and raw numbers are on [/benchmarks](https://corkcli.com/benchmarks). ## The CooperBench run we have planned CooperBench publishes its code (), its dataset and a leaderboard (), and its authors invite others to close the Solo–Coop gap. A CooperBench run with Cork is planned. When it is done, it goes on /benchmarks the way every study does, whatever it shows; [Study 8](https://corkcli.com/benchmarks/study-8-sixteen-agents-a-regression) is there because a build made things worse. Until then, we make no claim about how Cork does on CooperBench. ## Takeaways - CooperBench (arXiv:2601.13295) finds coding agents do worse together than one agent doing both tasks: 30% lower success on average in the abstract, 41% of Solo capability lost pooled in the appendix, up to about 50% for the strongest models. - Adding a chat channel did not significantly help; the largest share of the failures the authors reviewed (42%) came from agents' wrong picture of what their partner was doing. - Cork is built to give agents that picture from the code itself. Whether that closes the gap is untested; a CooperBench run is planned, and we will publish what it shows. ## Sources Checked October 19, 2026: - Khatua et al., "CooperBench: Why Coding Agents Cannot be Your Teammates Yet", arXiv:2601.13295 (v1 January 19, 2026; v2 January 26, 2026): ; abstract, sections 1, 2, 4, 5, 6 and appendix C in - CooperBench site and leaderboard: , - CooperBench code: - Cork benchmarks: ## Related reading - [How to run Codex and Claude Code side by side on one project](https://corkcli.com/blog/run-codex-and-claude-code-side-by-side): Run OpenAI Codex CLI and Claude Code on one repo: one AGENTS.md for both, a worktree each, shared MCP servers, matching hooks, and live coordination. - [Agents drifted into a language outsiders can't follow: what the Emergence World study found](https://corkcli.com/blog/emergence-world-study-agent-language): Emergence AI's Emergence World study found every agent world developed shared vocabulary outsiders struggle to follow. What the paper measured, what it did not, and how Cork keeps agent messages readable. - [Cork vs GitHub Agent HQ: agents side by side in isolation, or coordinated on one codebase](https://corkcli.com/blog/cork-vs-github-agent-hq): Cork vs GitHub Agent HQ: Agent HQ runs Copilot, Claude and Codex in GitHub's cloud, each task returning its own pull request; Cork coordinates agents working on the same code at once. - [Cork vs Superconductor: a hosted agent workspace, or coordination for the agents you run](https://corkcli.com/blog/cork-vs-superconductor): Cork vs Superconductor: Superconductor runs coding agents in cloud sandboxes for teams. Cork stops parallel agents from overwriting each other's code. [All articles](https://corkcli.com/blog.md) · [Cork in Markdown](https://corkcli.com/index.md) · [llms.txt](https://corkcli.com/llms.txt) This page in HTML: https://corkcli.com/blog/cooperbench-curse-of-coordination