Agent Orchestrator
A control plane for coordinating multiple LLM agents with budgets, guardrails, and human-in-the-loop approvals.
- Python
- FastAPI
- PostgreSQL
- LangGraph
- Docker
The Problem
Multi-agent systems drift, loop, and overspend. Production agent workloads need the same controls as any distributed system: budgets, timeouts, and audit trails.
Architecture
API Gateway ──▶ Orchestrator ──▶ Agent Pool
│ │
▼ ▼
Run Ledger Tool Sandbox
Key decisions:
- Token budgets per run — hard caps enforced at the orchestrator, not the agent.
- Approval gates — destructive tool calls pause for human confirmation.
- Run ledger — every agent step recorded for replay and cost attribution.
Lessons Learned
- Agents need fewer tools with better descriptions, not more tools.
- Deterministic orchestration beats emergent behavior for business workflows.
Placeholder writeup — replace with the real project deep-dive before launch.