Codex caps subagents at 3 by default; one config line unlocks 64

danielmac8 discovered that OpenAI Codex caps concurrent subagents at 3 by default, while OpenAI's internal Astra testing dispatched up to 64 Luna subagents in parallel. Raising the cap requires only a maxconcurrent setting under [agents] in config.toml. The finding was widely reshared the same day.

Confirmed

• Codex defaults to 3 concurrent subagents, adjustable via the [agents] maxconcurrent config in config.toml, up to 64
• OpenAI's internal Astra tests reportedly used up to 64 parallel Luna subagents (per danielmac8)
• After maxing concurrency to 64, danielmac8 observed UI lag, suspecting local CPU overload from parallel tasks

Why it matters

• The default limit of 3 is clearly insufficient for complex orchestration workloads like Astra-style multi-agent flows; the config unlocks larger-scale parallelism for advanced users
• High concurrency is not free: local CPU load is a real bottleneck, so users should weigh parallel gains against machine load before copying the config
• altryne proposed an alternative: instead of tuning concurrency, spawn Codex sessions directly as subagents—easier to interact with and no count limit—though danielmac8 noted this bypasses the multi-agent v2 path and loses native inter-agent communication

2026-09-07 ~ 2026-09-07 · 5 related posts

• [source] Codex defaults to 3 subagents — one config.toml line raises concurrency to 64 — daniel_mac8 · 2026-09-07
• [source] Follow-up: running 64 concurrent Codex subagents makes the UI laggy, likely CPU overload — daniel_mac8 · 2026-09-07
• [source] Spinning up Codex sessions in Claude Code loses native inter-agent communication — daniel_mac8 · 2026-09-07

2 near-duplicate retellings: daniel_mac8 · altryne