What this video shows
Dziri presents Cohere's controlled study of lead-and-worker coding teams. On ProgramBench, where agents must discover a hidden executable's behavior, teams improved on a solo agent but usually cost three to seven times as much.
The strongest cost result used one capable lead and 16 cheaper planning workers. It came within 1.4 points of a strong five-agent team at 70% lower cost. On Commit0-Lite, implementation delegation performed best in eight of ten tested team setups. Cohere describes that smaller comparison as suggestive rather than conclusive.
Read Cohere's research summary and exact comparisons before applying the 70% figure to another workload. Inspect the paper and experiment design for the tested models, prompts and benchmark limits. Compare the tasks with the original ProgramBench paper and Commit0 paper.
What you will learn
- Measure a team against a capable solo agent before paying for parallel work.
- Put workers where independent investigation adds information that the lead can use.
- Keep final integration with a capable lead when cheaper workers gather evidence.
- Re-run the comparison when the task structure, model or tool harness changes.
How to apply this safely
- Choose twenty representative repository tasks and define pass criteria, time and cost limits.
- Run the same lead model alone and record accepted output, tool calls, elapsed time and token cost.
- Label the stage that limits each task: discovery, planning, implementation or verification.
- Add two workers only at that stage, require compact evidence reports and keep the lead responsible for integration.
- Increase workers only while repeated runs improve accepted output enough to justify cost and latency.
Important limitations
- Cohere conducted the study and supplied the models, harness and cost comparison.
- The 70% reduction compares one heterogeneous 17-agent setup with a strong homogeneous team, not with the solo baseline.
- ProgramBench rewards black-box behavioral discovery, which differs from ordinary maintenance in a known repository.
- Commit0-Lite contains only 16 tasks and was near saturation for the strongest tested model, so its stage comparison is directional.
- More workers can increase latency, duplicate work and create integration failures even when aggregate benchmark reward rises.
Sources to check
- How to Get Frontier Coding Results at 70% Lower Cost AI Engineer talk by Nouha Dziri, published November 7, 2026.
- Lessons from scaling teams of frontier coding agents: reach the frontier at 30% of the cost Cohere research summary published October 14, 2026.
- Scaling Inference-Time Compute with Teams of Coding Agents Research paper with methods, results and limitations.
- ProgramBench Paper defining the black-box program reconstruction benchmark.
- Commit0 Paper defining the library-generation benchmark.
Continue learning on Learnetto
Reduce coding agent cost
Measure retries, routing, caching and accepted work.
Build an AI agent test set
Create representative tasks and repeatable acceptance checks.
Review AI-generated code
Keep independent evidence and human ownership at the merge boundary.