Reusable multi-agent orchestration workflow

Codex issue draft: reusable multi-agent orchestration workflow

Problem

Large coding tasks often need one strong planning agent and many cheaper execution
agents. Today the user must repeatedly explain the orchestration pattern:

• use the best available model for planning, issue decomposition, and integration;
• break work into zero-context packets that cheaper workers can execute safely;
• keep worker write scopes disjoint;
• route worker questions back to the orchestrator, not directly to the user;
• run validation waves before final handoff.

This is repetitive and easy to get wrong, especially when workers start with no
context.

Proposed Codex capability

Add a built-in skill or plugin workflow named multi-agent-orchestrator that
turns a user goal into an executable parallel work graph.

The orchestrator should:

1. Inspect the repo and relevant AGENTS.md files before planning.
2. Produce a dependency graph of subtasks with explicit read/write ownership.
3. Generate self-contained worker packets:

• objective;
• repo path;
• files/directories owned by the worker;
• forbidden files/directories;
• required commands and tests;
• local project instructions;
• default decisions;
• question policy;
• final response format.

1. Spawn cheap worker agents for implementation, exploration, and validation.
2. Use the strongest model for planning, resolving worker questions, and final

integration.

1. Escalate model quality only when a worker fails twice, produces conflicting

evidence, or hits a design decision not covered by the packet.

1. Keep the user out of the loop unless the orchestrator cannot answer a product

or risk decision from local context.

Codex-specific fit

This should map cleanly onto existing Codex concepts:

• Skills: package the reusable orchestration instructions.
• Plugins: distribute the skill plus optional helper scripts/templates.
• AGENTS.md: inject repo-specific workflow and safety rules into packets.
• Subagents: use spawn_agent/worker roles for parallel tasks.
• Plan mode: produce the work graph before execution when required.
• Thread goals: keep the orchestrator anchored to the overall objective.

Worker question policy

Worker agents should not ask the user directly. They should:

1. Use the default in their task packet when safe.
2. Return QUESTION_FOR_ORCHESTRATOR with the exact missing decision when no

safe default exists.

1. Stop only for true blockers, not routine uncertainty.

The orchestrator decides whether the answer is discoverable locally, can be
defaulted, or must be asked of the user.

Acceptance criteria

• A user can invoke one reusable skill/command instead of pasting a long custom

prompt.

• Codex creates a parallelizable plan with disjoint worker ownership.
• Workers receive enough context to start from zero context.
• The orchestrator integrates worker results and runs validation.
• The workflow supports cheap-worker execution with strong-model planning.
• The final response includes changed files, tests, blockers, and unresolved

risks.