{"id":"7e109589e89588d2531c27d4eabe1ee26abcf68efed6973b0164c9173270251f","title":"Cohere reports that coding-agent teams work best when workers target the task’s bottleneck","argument":"Cohere’s article, published October 14, 2026, argues that how coding-agent teams divide work matters more than simply adding agents. On ProgramBench, agents rebuild programs by testing a working original they cannot read. Cohere reports that workers helped most with exploration and planning: they gathered information for a lead to use while implementing and checking the rebuilt program. Delegating planning alone matched the reward of delegating all stages at half the cost. A Sol lead with cheaper planning workers came within 2.7 points of an all-Sol team that delegated every stage, for a fifth of its cost. These results come from a controlled testbed; the article does not establish that teams will be more effective or economical in other settings. Its explanation is specific to the tasks: ProgramBench requires agents to discover the original program’s behavior, while Commit0 supplies specifications and favored implementation work. Cohere concludes that much of a team’s performance depended on the lead’s choices—what to delegate and how to use workers’ reports—and says that as models improve at leading, the best team organization may change. Whether newer models generally lead teams better remains unknown; Cohere says its comparison used one run per model.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on November 5, 2026; originally published October 14, 2026. The reported results come from Cohere’s coding-agent testbed and ProgramBench; the article describes Commit0 as a secondary check with only 16 tasks. Cohere reports that results for models other than GPT-5.6 used one run per model on a 30-task subset, and says differences of a few points may be noise.","speaker":{"name":"Cohere","roleAtSource":"Publisher"},"source":{"title":"Scaling agent teams: reach the frontier at 30% of the cost","publisher":"Cohere","format":"Article","url":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","publishedAt":"2026-10-14T16:48:00.000Z","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"Cohere reports that, on ProgramBench, asynchronous Sol teams delegating only planning achieved about the same reward as teams delegating all stages at half the cost.","excerpt":"a lead that delegates only planning matched the reward of one that delegates every stage (63.6% against 63.2%) at half the cost ($249 against $510).","url":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","relationship":"Source-stated"},{"text":"Cohere reports that a Sol lead with Luna planning workers came within 2.7 points of an all-Sol team delegating every stage for a fifth of the cost.","excerpt":"we came within 2.7 points of it for a fifth of the cost.","url":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","relationship":"Source-stated"},{"text":"Cohere says that implementation, rather than planning, was the most effective stage to delegate on Commit0, while noting that the evidence is suggestive rather than conclusive.","excerpt":"implementation, not planning, was the most effective stage to delegate in eight of our ten team setups. However, frontier models like GPT-5.6 Sol are close to saturating Commit0","url":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","relationship":"Source-stated"},{"text":"Cohere says it cannot determine from one run per model whether newer models are generally better at leading teams.","excerpt":"Whether newer models are generally better at leading a team is a question we cannot settle with one run per model.","url":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","relationship":"Source-stated"},{"text":"Cohere concludes that team performance depended substantially on the lead and that the best organization may change as models improve at leading.","excerpt":"Much of what made our teams work depended on the lead: what it chose to delegate, and what it did with the reports that came back. As models get better at leading, the best way to organize a team may change too.","url":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well do these findings carry over to coding tasks whose bottleneck is not discovering a program’s behavior?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Cohere reports that coding-agent teams work best when workers target the task’s bottleneck","summary":"Cohere’s article, published October 14, 2026, argues that how coding-agent teams divide work matters more than simply adding agents. On ProgramBench, agents rebuild programs by testing a working original they cannot read. Cohere reports that workers helped most with exploration and planning: they gathered information for a lead to use while implementing and checking the rebuilt program. Delegating planning alone matched the reward of delegating all stages at half the cost. A Sol lead with cheaper planning workers came within 2.7 points of an all-Sol team that delegated every stage, for a fifth of its cost. These results come from a controlled testbed; the article does not establish that teams will be more effective or economical in other settings. Its explanation is specific to the tasks: ProgramBench requires agents to discover the original program’s behavior, while Commit0 supplies specifications and favored implementation work. Cohere concludes that much of a team’s performance depended on the lead’s choices—what to delegate and how to use workers’ reports—and says that as models improve at leading, the best team organization may change. Whether newer models generally lead teams better remains unknown; Cohere says its comparison used one run per model.","question":"How well do these findings carry over to coding tasks whose bottleneck is not discovering a program’s behavior?","claims":[{"text":"Cohere reports that, on ProgramBench, asynchronous Sol teams delegating only planning achieved about the same reward as teams delegating all stages at half the cost.","excerpt":"a lead that delegates only planning matched the reward of one that delegates every stage (63.6% against 63.2%) at half the cost ($249 against $510).","support":"source-stated"},{"text":"Cohere reports that a Sol lead with Luna planning workers came within 2.7 points of an all-Sol team delegating every stage for a fifth of the cost.","excerpt":"we came within 2.7 points of it for a fifth of the cost.","support":"source-stated"},{"text":"Cohere says that implementation, rather than planning, was the most effective stage to delegate on Commit0, while noting that the evidence is suggestive rather than conclusive.","excerpt":"implementation, not planning, was the most effective stage to delegate in eight of our ten team setups. However, frontier models like GPT-5.6 Sol are close to saturating Commit0","support":"source-stated"},{"text":"Cohere says it cannot determine from one run per model whether newer models are generally better at leading teams.","excerpt":"Whether newer models are generally better at leading a team is a question we cannot settle with one run per model.","support":"source-stated"},{"text":"Cohere concludes that team performance depended substantially on the lead and that the best organization may change as models improve at leading.","excerpt":"Much of what made our teams work depended on the lead: what it chose to delegate, and what it did with the reports that came back. As models get better at leading, the best way to organize a team may change too.","support":"source-stated"}],"limitations":["Text as retrieved on November 5, 2026; originally published October 14, 2026.","The reported results come from Cohere’s coding-agent testbed and ProgramBench; the article describes Commit0 as a secondary check with only 16 tasks.","Cohere reports that results for models other than GPT-5.6 used one run per model on a 30-task subset, and says differences of a few points may be noise."],"mediaId":null,"correctionReason":null,"id":"7e109589e89588d2531c27d4eabe1ee26abcf68efed6973b0164c9173270251f","publicationId":"ab3dd05d-5714-b543-ac6c-7d376be78112","version":1,"sourceUrl":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","sourceTitle":"Scaling agent teams: reach the frontier at 30% of the cost","publisher":"Cohere","publishedAt":"2026-10-14T16:48:00.000Z","originalPublished":{"value":"2026-10-14T16:48:00.000Z","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-11-05T07:52:01.356Z","createdAt":"2026-11-06T00:03:24.407Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"dc8365c835ae4f00f1423b593133c69aff380640ea0ed0b3b2e7cfa6be344e6c","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-11-05T07:51:58.413Z"},"provenance":{"textHash":"ca4b4ad76ff584f24afe58c1ea5aca6f387e49c48f1895a557b9b9cafe782c7e","observationId":"d2739a01d0978e8bb75890d6ce7b5303a4178bae6f84976d99f9642a7e1b2fd0","retrievedAt":"2026-11-05T07:53:36.985Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"90fe191a-24ba-c088-ac6d-07734b78a101"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"8ea5565c-1837-6618-68ff-9bf27770c56f"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-05T23:58:41.826Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/cohere-reports-that-coding-agent-teams-work-best-when-workers-7e109589","formats":{"html":"/readings/cohere-reports-that-coding-agent-teams-work-best-when-workers-7e109589","markdown":"/readings/cohere-reports-that-coding-agent-teams-work-best-when-workers-7e109589.md","json":"/readings/cohere-reports-that-coding-agent-teams-work-best-when-workers-7e109589.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[],"entities":[]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"7e109589e89588d2531c27d4eabe1ee26abcf68efed6973b0164c9173270251f","revision":"ab3dd05d-5714-b543-ac6c-7d376be78112","revisionKind":"publication-id","title":"Cohere reports that coding-agent teams work best when workers target the task’s bottleneck","url":"https://escapemeridian.com/readings/cohere-reports-that-coding-agent-teams-work-best-when-workers-7e109589"},"source":{"type":"source","id":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","revision":"ca4b4ad76ff584f24afe58c1ea5aca6f387e49c48f1895a557b9b9cafe782c7e","revisionKind":"sha256-source-text","title":"Scaling agent teams: reach the frontier at 30% of the cost","url":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"7e109589e89588d2531c27d4eabe1ee26abcf68efed6973b0164c9173270251f","title":"Cohere reports that coding-agent teams work best when workers target the task’s bottleneck","argument":"Cohere’s article, published October 14, 2026, argues that how coding-agent teams divide work matters more than simply adding agents. On ProgramBench, agents rebuild programs by testing a working original they cannot read. Cohere reports that workers helped most with exploration and planning: they gathered information for a lead to use while implementing and checking the rebuilt program. Delegating planning alone matched the reward of delegating all stages at half the cost. A Sol lead with cheaper planning workers came within 2.7 points of an all-Sol team that delegated every stage, for a fifth of its cost. These results come from a controlled testbed; the article does not establish that teams will be more effective or economical in other settings. Its explanation is specific to the tasks: ProgramBench requires agents to discover the original program’s behavior, while Commit0 supplies specifications and favored implementation work. Cohere concludes that much of a team’s performance depended on the lead’s choices—what to delegate and how to use workers’ reports—and says that as models improve at leading, the best team organization may change. Whether newer models generally lead teams better remains unknown; Cohere says its comparison used one run per model.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on November 5, 2026; originally published October 14, 2026. The reported results come from Cohere’s coding-agent testbed and ProgramBench; the article describes Commit0 as a secondary check with only 16 tasks. Cohere reports that results for models other than GPT-5.6 used one run per model on a 30-task subset, and says differences of a few points may be noise.","speaker":{"name":"Cohere","roleAtSource":"Publisher"},"source":{"title":"Scaling agent teams: reach the frontier at 30% of the cost","publisher":"Cohere","format":"Article","url":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","publishedAt":"2026-10-14T16:48:00.000Z","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"Cohere reports that, on ProgramBench, asynchronous Sol teams delegating only planning achieved about the same reward as teams delegating all stages at half the cost.","excerpt":"a lead that delegates only planning matched the reward of one that delegates every stage (63.6% against 63.2%) at half the cost ($249 against $510).","url":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","relationship":"Source-stated"},{"text":"Cohere reports that a Sol lead with Luna planning workers came within 2.7 points of an all-Sol team delegating every stage for a fifth of the cost.","excerpt":"we came within 2.7 points of it for a fifth of the cost.","url":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","relationship":"Source-stated"},{"text":"Cohere says that implementation, rather than planning, was the most effective stage to delegate on Commit0, while noting that the evidence is suggestive rather than conclusive.","excerpt":"implementation, not planning, was the most effective stage to delegate in eight of our ten team setups. However, frontier models like GPT-5.6 Sol are close to saturating Commit0","url":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","relationship":"Source-stated"},{"text":"Cohere says it cannot determine from one run per model whether newer models are generally better at leading teams.","excerpt":"Whether newer models are generally better at leading a team is a question we cannot settle with one run per model.","url":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","relationship":"Source-stated"},{"text":"Cohere concludes that team performance depended substantially on the lead and that the best organization may change as models improve at leading.","excerpt":"Much of what made our teams work depended on the lead: what it chose to delegate, and what it did with the reports that came back. As models get better at leading, the best way to organize a team may change too.","url":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well do these findings carry over to coding tasks whose bottleneck is not discovering a program’s behavior?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Cohere reports that coding-agent teams work best when workers target the task’s bottleneck","summary":"Cohere’s article, published October 14, 2026, argues that how coding-agent teams divide work matters more than simply adding agents. On ProgramBench, agents rebuild programs by testing a working original they cannot read. Cohere reports that workers helped most with exploration and planning: they gathered information for a lead to use while implementing and checking the rebuilt program. Delegating planning alone matched the reward of delegating all stages at half the cost. A Sol lead with cheaper planning workers came within 2.7 points of an all-Sol team that delegated every stage, for a fifth of its cost. These results come from a controlled testbed; the article does not establish that teams will be more effective or economical in other settings. Its explanation is specific to the tasks: ProgramBench requires agents to discover the original program’s behavior, while Commit0 supplies specifications and favored implementation work. Cohere concludes that much of a team’s performance depended on the lead’s choices—what to delegate and how to use workers’ reports—and says that as models improve at leading, the best team organization may change. Whether newer models generally lead teams better remains unknown; Cohere says its comparison used one run per model.","question":"How well do these findings carry over to coding tasks whose bottleneck is not discovering a program’s behavior?","claims":[{"text":"Cohere reports that, on ProgramBench, asynchronous Sol teams delegating only planning achieved about the same reward as teams delegating all stages at half the cost.","excerpt":"a lead that delegates only planning matched the reward of one that delegates every stage (63.6% against 63.2%) at half the cost ($249 against $510).","support":"source-stated"},{"text":"Cohere reports that a Sol lead with Luna planning workers came within 2.7 points of an all-Sol team delegating every stage for a fifth of the cost.","excerpt":"we came within 2.7 points of it for a fifth of the cost.","support":"source-stated"},{"text":"Cohere says that implementation, rather than planning, was the most effective stage to delegate on Commit0, while noting that the evidence is suggestive rather than conclusive.","excerpt":"implementation, not planning, was the most effective stage to delegate in eight of our ten team setups. However, frontier models like GPT-5.6 Sol are close to saturating Commit0","support":"source-stated"},{"text":"Cohere says it cannot determine from one run per model whether newer models are generally better at leading teams.","excerpt":"Whether newer models are generally better at leading a team is a question we cannot settle with one run per model.","support":"source-stated"},{"text":"Cohere concludes that team performance depended substantially on the lead and that the best organization may change as models improve at leading.","excerpt":"Much of what made our teams work depended on the lead: what it chose to delegate, and what it did with the reports that came back. As models get better at leading, the best way to organize a team may change too.","support":"source-stated"}],"limitations":["Text as retrieved on November 5, 2026; originally published October 14, 2026.","The reported results come from Cohere’s coding-agent testbed and ProgramBench; the article describes Commit0 as a secondary check with only 16 tasks.","Cohere reports that results for models other than GPT-5.6 used one run per model on a 30-task subset, and says differences of a few points may be noise."],"mediaId":null,"correctionReason":null,"id":"7e109589e89588d2531c27d4eabe1ee26abcf68efed6973b0164c9173270251f","publicationId":"ab3dd05d-5714-b543-ac6c-7d376be78112","version":1,"sourceUrl":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","sourceTitle":"Scaling agent teams: reach the frontier at 30% of the cost","publisher":"Cohere","publishedAt":"2026-10-14T16:48:00.000Z","originalPublished":{"value":"2026-10-14T16:48:00.000Z","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-11-05T07:52:01.356Z","createdAt":"2026-11-06T00:03:24.407Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"dc8365c835ae4f00f1423b593133c69aff380640ea0ed0b3b2e7cfa6be344e6c","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-11-05T07:51:58.413Z"},"provenance":{"textHash":"ca4b4ad76ff584f24afe58c1ea5aca6f387e49c48f1895a557b9b9cafe782c7e","observationId":"d2739a01d0978e8bb75890d6ce7b5303a4178bae6f84976d99f9642a7e1b2fd0","retrievedAt":"2026-11-05T07:53:36.985Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"90fe191a-24ba-c088-ac6d-07734b78a101"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"8ea5565c-1837-6618-68ff-9bf27770c56f"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-05T23:58:41.826Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"ab3dd05d-5714-b543-ac6c-7d376be78112","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading attributes the testbed, results, explanations and conclusions to Cohere, the publisher of this single blog article. Findings from other work, such as Factory and Meta, are not presented as Cohere's own. Claims stay attributed rather than stated as established fact.","excerpt":"To answer it, we built a controlled testbed and ran teams of coding agents on it."},{"dimension":"summary","status":"uncertain","explanation":"The summary is faithful, but central terms are unexplained and no definitions are supplied. Sol and Luna are never identified as GPT-5.6 models ranked by capability and cost, so a general reader may assume they are Cohere models. Reward is not described as the share of hidden behavioral tests passed, which leaves '2.7 points' unclear. The lead/worker setup is also only implied. Smallest repair: one clause identifying Sol and Luna as stronger and cheaper GPT-5.6 models, plus a brief gloss of reward.","excerpt":"We focus primarily on three GPT-5.6 models, Luna, Terra, and Sol, in increasing order of capability and cost."},{"dimension":"date","status":"pass","explanation":"The stated date, October 14, 2026, matches the UTC calendar day of originalPublished. modifiedAt equals the publication instant, so there is no sign of a later edit. The evidence is a publisher page, not an archive copy, so no archive disclosure is needed. There is no hindsight framing, and forecasts in the article are not presented as outcomes.","excerpt":null},{"dimension":"support","status":"pass","explanation":"Each claim matches the source. Planning-only delegation matched all-stage delegation (63.6% vs 63.2%) at half the cost. The Luna-worker team came within 2.7 points for a fifth of the cost. Commit0 favored implementation and is treated as suggestive. The one-run-per-model caveat and the lead-dependence conclusion are both stated. The caveat about other settings and the question about non-discovery bottlenecks follow from the source's own discussion and imply no unsupported benefit.","excerpt":"Planning paid off on ProgramBench because the specification has to be discovered, but on other tasks the bottleneck may be elsewhere."},{"dimension":"media","status":"pass","explanation":"No media was selected and no verified candidates were supplied, so there is nothing to misidentify.","excerpt":null}],"createdAt":"2026-11-05T23:58:41.826Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}