Project 5: Multi-Agent Orchestration Sandbox
Orchestrate specialist agents with deterministic handoffs and traceable decisions.
Quick Reference
| Attribute | Value |
|---|---|
| Difficulty | Level 3 (Advanced) |
| Time Estimate | 1-2 weeks |
| Main Programming Language | TypeScript |
| Alternative Programming Languages | Python, Go |
| Key Topics | Agent roles, terminal states, traceability |
1. Learning Objectives
- Define strict role boundaries for planner and specialists.
- Implement terminal-state control to prevent loops.
- Capture branch rationale in structured traces.
2. All Theory Needed (Per-Concept Breakdown)
Concept A: Agent Topology Design
- Keep planner responsibilities separate from commit responsibilities.
- Start with minimal agent count.
Concept B: Branch Rationale Telemetry
- Every branch choice emits a reason code.
- Trace contains selected tools, retries, and final status.
3. Architecture and Build Plan
- Define two-agent baseline.
- Add one policy specialist for high-risk decisions.
- Set hop-count ceiling and terminal states.
4. Validation and Testing
- No traces exceed hop limit.
- Commit actions are never executed by non-commit agents.
- Branch decisions are explainable via reason codes.
5. Troubleshooting
- Symptom: planner-agent ping-pong.
- Fix: enforce terminal-state and ownership constraints.
6. Deliverables
- Agent responsibility matrix.
- Trace schema and sample traces.
- Loop-prevention test output.
7. Stretch Goals
- Add dynamic policy checks based on user risk profile.