Skip to content

Supervisory Team Workflow Contract

Minimal contract for the first user-shell supervisory experiment.

Purpose

Run one Codex session as a supervisor of a small native subagent team and learn how much useful coordination Codex can perform before any custom orchestration layer is added.

Supervisor Role

The supervisor is responsible for:

  • reading the task brief,
  • deciding what workstreams exist,
  • launching native subagents for substantive work,
  • clearly orienting subagents toward distinct workstreams,
  • comparing and synthesizing subagent results,
  • deciding what follow-on delegation is needed,
  • and returning a clear terminal result for the human operator.

Core Constraint

The supervisor should not do substantive task work itself.

For this experiment, substantive work includes:

  • writing or revising the main design content,
  • performing detailed critique that could have been delegated,
  • drafting implementation changes,
  • and carrying out direct analysis that belongs to a subagent workstream.

The supervisor may still do lightweight coordination work itself, such as:

  • reading the task brief,
  • reading high-level repo instructions,
  • selecting which subagents to launch,
  • comparing subagent outputs,
  • synthesizing a final recommendation,
  • and deciding when to stop.

Delegation Rule

All meaningful task elements should be delegated to subagents.

Examples:

  • if the task is design critique, the supervisor should assign critique angles to subagents
  • if the task is planning, the supervisor should assign candidate approaches or analysis passes to subagents
  • if the task is implementation shaping, the supervisor should assign codebase reading and option generation to subagents
  • if the task is review, the supervisor should assign file or file-set review purposes such as issues, consistency, summary, or comparative assessment

The supervisor may refine or sequence those delegations, but it should not collapse the experiment back into single-agent execution.

Allowed Native Behaviors

The supervisor may:

  • spawn one or more native subagents,
  • wait for subagent results,
  • send follow-up work to subagents if needed,
  • ask subagents to review files or file sets for a distinct purpose,
  • compare competing results,
  • and stop early if the task is blocked or the results are converging poorly.

Delegation Budget

Keep the experiment small.

Unless the task brief says otherwise, use no more than 5 subagent calls total.

The goal is to test supervisory leverage, not to maximize fan-out.

Stop Conditions

The supervisor should stop when one of the following is true:

  • the task brief has been addressed well enough for human review,
  • subagent results are repetitive or low-value,
  • the session is blocked on missing context or permissions,
  • or the time and effort budget in the task brief has been reached.

Required Final Output

The final response should be useful to a human reviewer.

It should include:

  • a short statement of outcome,
  • what subagent workstreams were used,
  • the main result or recommendation,
  • and any clear blockers or next steps.

What This Experiment Is Testing

This contract is not trying to prove that Codex can solve the task well.

It is trying to prove whether a shell-launched supervisor using native subagents can:

  • maintain a supervisory role,
  • delegate real work instead of absorbing it,
  • produce a useful synthesized result,
  • and do so in a way that appears stronger than a comparable straight-line single-agent pass.