Cohere Labs: teams of coding agents beat solo agents on ProgramBench, and a strong lead with cheap planning workers gets near-frontier results at about 30% of the cost

Why it matters

【Key points】

Cohere Labs published a study on Oct 14 on how to organize teams of coding agents. The main testbed is ProgramBench: 200 programs (from jq and ripgrep to SQLite, FFmpeg and PHP) that agents must rebuild from scratch by running the original binary and reading docs, scored by hidden behavioral tests. It mainly uses three GPT-5.6 tiers, Luna, Terra and Sol, three runs per setup.

With the same minimal prompt, a Sol team passed 61.8% vs 52.1% solo, at 4–7x the cost. Delegating only exploration and planning scored 63.6%, matching full delegation (63.2%) at half the cost ($249 vs $510). A Sol lead with cheap Luna planning workers gets near the all-Sol team at about 30% of the cost (key takeaways: 61.8% vs 63.2%, $510 to $155). Cheap workers cannot rescue a weak lead: Terra lead plus Luna workers hit 51.6%, below solo Sol's 57.7%. A direct messaging tool barely helped.

On a 30-task subset (single runs), Claude Opus 4.8 teams gained up to 5.9 points; Opus 4.7 rarely delegated; Claude Haiku 4.5 and Gemini 3.7 Flash delegated but gained nothing. Claude Opus 5 hit the 8-hour timeout on all eight trial tasks at about $500 each, roughly 15x an Opus 4.8 team, and was stopped.

【Verifiable】

Cohere blog: https://cohere.com/blog/lessons-from-scaling-teams-of-frontier-coding-agents

【Our read】

1. Where you add agents matters more than how many: on black-box rebuilds the bottleneck is discovering the spec, so planning pays most; on Commit0, where the spec is given, implementation pays more.

2. Limits: one harness and mainly one benchmark; other models ran once, so few-point gaps may be noise; Opus 5 timed out rather than failed.

3. Strong-lead, cheap-worker teams are a clear cost lever, and leading a team looks like a trainable skill that varies widely across models.