Subagents parallelize work — and they complicate approval. The approval model that worked for one agent breaks for several, and the fix is designing it before you scale.
Codex subagents let you fan out work to multiple agents in parallel. The productivity win is real; the governance question is who approves what, when there are several agents acting at once.
The Power Dynamic: Approval Across Agents
Approval for one agent is simple: a human reviews the action. Approval across subagents is harder — the actions interleave, the dependencies cross agents, and the human can't review every action without becoming the bottleneck. The model has to scale with the subagents.
- One agent: review every action.
- A few subagents: review by type, not by action.
- Many subagents: review by policy, with exceptions escalated.
The Conduct Rule: Approve by Risk, Not by Count
The approval model should be risk-tiered, not uniform. Low-risk, reversible actions run without approval; high-risk, irreversible actions require it. The rule of conduct is what's approved automatically and what isn't — written before the subagents run, not after the first incident.
- Low-risk, reversible: no approval.
- High-risk, irreversible: approval required.
- Exceptions: escalated with context.
What Breaks With Subagents
- Every action requires approval, so the human becomes the bottleneck and people bypass the gate.
- No action requires approval, so the subagents ship things nobody authorized.
- The approval model doesn't account for actions that cross subagents.
How to Design It
- Tier actions by risk and reversibility.
- Approve low-risk automatically; gate high-risk.
- Escalate exceptions with the context the human needs.
- Log every action so the parallel work is reconstructable.
Conclusion
Codex subagents parallelize work and complicate approval. The approval model has to scale with the subagents: risk-tiered, not uniform, with exceptions escalated and every action logged. The model is designed before you scale — or it's improvised after the first incident.
About FACTA
FACTA helps startups and growth-stage teams turn AI into production systems that keep running — not demos that impress once.
We design the architecture around the parts that actually break under real usage: tooling you own, credentials you control, failover, cost controls, observability. The boring infrastructure that keeps a system alive after launch.
Led by Matías Baglieri and Carolina Fogliato, we focus on one thing:
AI leadership that builds. Not just advises.
Tell us how you approve your Codex subagents today.
We'll tell you how to tier it by risk without becoming the bottleneck. See sandbox and approval for the sandbox side.
Explore AI Automation