Multi-Agent Operator Guide

Run a strict two-role architecture:

• One orchestrator for planning, locking, assignment, and integration.
• N parallel subagents for scoped execution only.

Load the role prompts from:

• references/ORCHESTRATOR-AGENT.md
• references/PARALLELIZABLE-SUBAGENT.md

Quick Start

1. Load references/ORCHESTRATOR-AGENT.md and instantiate exactly one orchestrator.
2. Provide objective, constraints, environment state, and available models.
3. Require orchestrator output in its mandatory schema.
4. Spawn N subagents (N < 10) with prompts generated by orchestrator.
5. Enforce lock scopes before execution starts.
6. Collect subagent outputs, validate acceptance checks, and merge via orchestrator.

Mandatory Operator Policy

• Never run subagents without explicit lock scopes.
• Never allow overlapping lock scopes across active subagents.
• Never allow a subagent to define its own task outside assignment.
• Always require append-only worklogs.
• Always route completion and retry decisions through orchestrator.

Recommended Model Selection

• Orchestrator: highest-capability model available.
• Subagents: smallest model that still meets acceptance criteria for assigned task.
• Escalate a subagent model only after a concrete failure.

Required Artifacts Per Run

• Canonical task ledger.
• Active lock table.
• Subagent assignment packets.
• Per-subagent worklog file.
• Final orchestrator integration report.

Troubleshooting

• Lock conflict: pause conflicting tasks, re-issue lock table, reassign.
• Missing context: mark task blocked and provide explicit missing inputs.
• Repeated failure: retry once with narrowed context, then replan serially.
• Scope drift: reject output and rerun with stricter forbidden-scope list.

Do Not

• Do not mix orchestration and execution duties in one agent.
• Do not pass full repository context to every subagent.
• Do not mark tasks done without criterion-by-criterion acceptance checks.