collaborative-agent

Testing collaborative agents coordinated by a strict orchestrator. Includes role separation, retry logic, and standardized output generation.

Usage

• Configure your LLM engine to handle at least 2 parallel requests.
• Use an agentic setup that can 1) launch a sub agent, 2) support autonomous (AKA YOLO) mode, and 3) read AGENTS.md at startup.
• Configure your agentic CLI to use your local LLM engine.
• Launch it from the repository root directory.
• Kick it off with an explicit instruction like: "Act as the orchestrator agent and perform the task."

If you are using Codex, update to the latest version and enable the collaborative agents by adding the following to ~/.codex/toml.

[features]
multi_agent = true

How it works

• AGENTS.md is the entry point.
• The orchestrator reads ORCHESTRATOR.md, then processes files in transcripts/ by spawning one worker per file (sequentially), reviews results, retries if needed, and writes task.md.
• Each worker reads WORKER.md, processes exactly one assigned transcript, and writes result/<same_name>.md.

Expected output

• result/ contains one markdown summary per transcript.
• task.md summarizes what was completed.