Which agent teams work on code?
Partitioning by code dependencies beats headcount: the Co-Coder measurement and lessons from Cursor, Cognition and Anthropic
Slide contents
1. Which agent teams work on code?
Partitioning by code dependencies beats headcount: the Co-Coder measurement and lessons from Cursor, Cognition and Anthropic
2. The fastest team writes the worst code
CodeProjectEval: 18 projects, one gpt-5-mini model for every setup
3. A team needs a code dependency map
Cost of links → partition → schedule → measurement and practice
4. We know a team can be worse
The last deck asked “when”; this one asks “how to split code”
5. Peer agents queue at one lock
Shared state file and locks: 20 agents work like 1–3
6. One big task does not split itself
16 agents, the Linux kernel and one shared bug
7. One agent can replay a uniform team
If all roles share one model, they can take turns
8. Team time is path plus links
T(P) = W(P) + α·C(P): critical path and context transfer
9. Interface blueprint first, agents second
An LLM writes the RIB from requirements; a judge checks up to 3 rounds
10. T2's shared module is the filter schema
A teaching split: API, UI and one shared schema
11. Shared files leave the groups
Utilities stand alone; aggregators share one integration group
12. One agent keeps coupled files together
Infomap picks the group count; leaves are lifted for parallelism
13. A file starts once its dependencies finish
One shared task list, no barriers between layers
14. Repairs go to the owning group
The leader runs tests and routes failures, up to 10 rounds
15. Every setup got the same blueprint
28 projects, gpt-5-mini, three runs, three metrics
16. Small projects show a smaller gain
DevEval: +11.3 pp, time 0.55×, cost 0.72×
17. Partitioning improves all three metrics
CodeProjectEval: agent per file costs more, Agent Teams loses quality
18. Denser links, more tests gained
Edge density ≈ 1.5 versus ≈ 0.3; r = 0.65 across 16 projects
19. One writes, the others think
A reviewer's clean context catches what the author misses
20. Hierarchy survived; the integrator did not
Planners split; isolated workers return one handoff
21. The preprint measures only from-scratch Python projects
28 tasks, one model, a graph from a blueprint, not live code
22. Split code by dependencies, not evenly
A team must earn its coordination cost