codex - 💡(How to fix) Fix Codex ignores explicit delegation-only instructions and duplicates delegated research in parent thread [5 comments, 4 participants]

Summary

On April 17, 2026, in thread 019d9aa2-7ff7-7e30-848c-bfe8f70dd8a1, Codex did not reliably respect explicit user instructions about delegation boundaries.

When I explicitly asked it to delegate research to a subagent so the parent thread would not bloat its context, it launched the agent and then still performed the same research itself in the parent thread.

Earlier in the same session, a local workflow/skill expected a checker/context-checker agent run, but Codex claimed a higher-priority rule prevented that. That restriction was not visible in the local skill files, repo AGENTS.md, or my visible ~/.codex/config.toml.

This makes multi-agent orchestration feel unreliable and defeats the purpose of delegation.

Environment

• Observed date: April 17, 2026
• Thread uploaded via /feedback: 019d9aa2-7ff7-7e30-848c-bfe8f70dd8a1
• Local Codex version cache showed latest version 0.121.0 at the time of filing
• Running in a Codex terminal/agent workflow environment with subagent support

Repro

1. Use a workflow or skill file that explicitly expects a subagent launch.
2. Ask Codex to investigate something using a delegated research agent because you do not want the parent thread to accumulate that research context.
3. Clarify if needed that by “search it up” you specifically mean “delegate an agent to do the research.”
4. Observe that Codex may still do local/web research itself in the parent thread after launching the agent.
5. Also observe that Codex may claim a higher-priority restriction on subagent use that is not visible from normal local repo/config surfaces.

Expected behavior

• If the user explicitly requests agent-only delegation, the parent should not also perform the same research unless the user asked for parallel verification.
• If a higher-priority instruction conflicts with local workflow/skill instructions, Codex should surface that conflict before acting.
• Delegation policy that materially changes behavior should be auditable from normal config or session surfaces.

Actual behavior

• Delegation boundaries were not followed tightly.
• Parent-thread research duplicated work that had already been delegated.
• A hidden or non-obvious policy appeared to override visible workflow intent.

Why this matters

• It causes unnecessary context bloat in exactly the scenarios where subagents are supposed to prevent it.
• It wastes tokens and reduces trust in xhigh orchestration.
• It makes local workflow systems/skills unreliable when visible instructions can be silently overridden by non-obvious policy layers.

Additional note

This report is about product behavior, not just prompt quality. The problem is specifically that explicit user constraints on delegation strategy were not obeyed reliably, and any conflicting higher-priority policy was not made auditable from the normal visible configuration surfaces.