Summary: Cohere published results on October 14, 2026 from a controlled testbed where teams of coding agents rebuild programs from scratch, and the findings reshape how you should spend inference compute. Well-organized agent teams beat solo agents by up to roughly 10 percentage points on ProgramBench, but cost 3–7x more — unless you allocate the agents where they matter. The headline result: a strong lead model with cheap workers reaches near-frontier performance at about 30% of the cost of an all-expensive team.
What Cohere actually tested
Cohere built a deliberately minimal harness, inspired by mini-SWE-agent, where one lead agent receives a task and delegates subtasks to workers through a shared bash workspace. Workers see only their subtask and report back; each delegation either spawns a fresh worker or resumes one that already has context, so the lead can ask the same worker to fix its own bugs rather than brief a new one (Cohere Blog).
They varied five dimensions: how much guidance the lead gets (a minimal prompt versus a structured explore–plan–implement–verify workflow), whether the lead blocks or works asynchronously, whether workers run the same model as the lead or a cheaper one, team size from two to sixteen workers, and whether agents can message each other directly. The benchmark is ProgramBench: 200 programs, from small tools like jq to systems like SQLite and FFmpeg, where the agent can run but not read the original program and must reproduce its behavior. Each configuration ran three times across all 200 tasks, with per-task comparisons against a solo baseline.
Teams win, but the bill multiplies
The baseline comparison is blunt. A blocking Sol team passed 61.8% of hidden behavioral tests on average versus 52.1% for a solo Sol agent, and the four Sol team configurations cost 4–7x the corresponding solo run. Factory’s own three-role system, cited by Cohere as context, more than doubled reward on a GDAL rebuild but at 14x the cost and 13x the duration.
So the interesting question shifts from “do teams help” to “where does each extra agent earn its keep.” Cohere’s answer is specific.
Allocation beats headcount
When the lead delegated only one stage and handled the rest itself, planning was the clear winner. A team delegating only exploration and planning scored 63.6%, statistically matching the 63.2% of a team delegating every stage — at half the cost (USD 249 versus USD 510). Measured as reward gained per USD 100 of added cost, planning was about twice as efficient as the next-best stage. The reason is structural: the target program is a black box, so planning workers can probe different parts of its interface in parallel and hand the lead a fuller picture before it writes any code.\n Asynchrony cut wall-clock time 19.5% with the structured prompt at equal reward, though it had a trap: with the minimal prompt, the async lead hoarded work, delegated less often, and scored below even a solo agent given the structured prompt.
The 70%-cheaper team
The most practical result is the model mix. A Sol lead coordinating sixteen inexpensive Luna planning workers reached 61.8% reward, versus 63.2% for an all-Sol team delegating every stage — a 1.4-point gap that cut cost from USD 510 to USD 155. That is a senior-engineer-plus-juniors pattern, and it held up in a controlled setting.
If you’re sizing an agent stack by request shape rather than raw volume — the same discipline I argued for when sizing embedding and rerank infrastructure — this is the agentic equivalent: match model tier to role, and match delegation to the bottleneck stage, not to everything.
What to take, and what to check
Two caveats before you redesign your pipeline. First, ProgramBench rewards discovery because the spec is hidden; on Commit0-Lite, where documentation and tests are given, Cohere used it only as a secondary check, and the supplied post doesn’t detail whether planning delegation transfers when the specification is already explicit. Second, real-world team projects cited in the post — sixteen Claude agents writing a 100,000-line C compiler, Cursor’s agents building a browser — used very different structures (peer lock files, recursive hierarchy, fixed roles), so Cohere’s numbers describe their harness, not those systems.
The actionable takeaway: before adding agents, decide which stage actually binds your task, spend your expensive model on the lead role only, and instrument reward-per-dollar rather than reward alone. Then test whether async delegation is helping or just letting your lead avoid delegating.