Human attention becomes the concurrency ceiling
Source → Observed → Interpretation → Model implication
Open-source Codex orchestration with Symphony
View source →OpenAI reports that engineers could comfortably supervise only three to five Codex sessions at once; beyond that productivity dropped as engineers lost track of work and debugged stalled agents.
Symphony moves coordination to an always-on orchestrator where open tasks receive agents in isolated workspaces and humans review outcomes.
OpenAI reports a 500% increase in landed pull requests on some teams.
Cooperation can begin before rich agent-to-agent collaboration. Independent agents first create a coordination problem; explicit orchestration then allows their work to compose as a system.
REFINES. Refines the Selection → Cooperation boundary: coordination infrastructure may be a mechanism through which Selection produces Cooperation.
What this does not establish
- It does not establish operation of 1,000 concurrent coding agents.
- It does not establish stable Specialization among agents.
- The reported productivity increase applies to some teams and should not be generalized to OpenAI as a whole.
Does coordination infrastructure consistently allow useful concurrency to grow beyond the human supervision ceiling without coordination cost becoming dominant?