Multi-agent orchestration for GitHub Copilot CLI — autonomous loops, parallel tmux agent teams, and durable scheduled jobs.
💡 Patterns (Architecture & Execution)
- Autonomous Agentic Loops: Agents follow a strict Reason → Act → Observe cycle. They autonomously execute tools, parse stdout/stderr, and self-correct dynamically without user intervention.
- Parallel Tmux Teams: Tmux multiplexing serves as the spatial orchestration layer. Specialized agents (e.g., architect, debugger, executor) run concurrently in isolated terminal panes.
- Durable Stateful Jobs: Background execution utilizes durable task frameworks rather than ephemeral cron schedules. State and episodic memory survive process crashes and network disconnects.
🚨 Anti-patterns (Gotchas & Traps)
- Monolithic Context Windows: Cramming multiple agent personas into a single context stream creates severe performance bottlenecks and context pollution. Isolate processes via Tmux.
- Fire-and-Forget Cron: Standard cron jobs lose execution state upon failure. Agents will blindly restart from scratch instead of resuming mid-task.
- Yielding on Failure: Chatbots stop when an error occurs. Autonomous loops must trap the error, analyze the stack trace, and mutate the approach instead of halting.
🚀 Best Practices (Implementation & Reliability)
- Episodic Memory Management: Persistently store daily task context and execution telemetry. This enables agents to identify long-term trends and resume complex multi-day projects seamlessly.
- Role-Based Pane Assignment: Dedicate specific Tmux windows to distinct specialist personas to isolate crash domains. One pane runs implementation, another monitors telemetry, a third executes continuous tests.
- Continuous Feedback Ingestion: Pipe raw test suites and linter outputs directly back into the agent’s context window. The autonomous loop requires immediate, unadulterated failure data to adjust the next iteration effectively.