Senior GenAI Engineer — Fix Unstable Multi-Agent System (Urgent)

We have a multi-agent LLM system in production that has become unreliable, and we need a true expert to stabilize it — not someone learning agents on our budget. Current symptoms: • Agents loop, stall, or hand off to the wrong agent • Non-deterministic behavior — same input, different outcome • Tool/function calls fail silently or return malformed output • Context bloat and runaway token cost • No reliable way to trace what went wrong in a run What we need you to do: 1. Audit the existing orchestration, prompts, tool schemas, and state handling 2. Deliver a written diagnosis of root causes (not guesses) 3. Re-architect or repair the agent graph: routing, handoffs, retries, termination conditions 4. Add observability — tracing, run logs, evals so failures are reproducible 5. Harden it: guardrails, fallbacks, cost controls, load testing You must have: • 10+ years in software/AI engineering, with deep recent hands-on multi-agent work • Production experience with LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, or a custom orchestrator you built yourself • Strong Python, async, and distributed-systems debugging skills • Real experience with RAG, vector DBs, tool/function calling, and structured output enforcement • Evaluation and observability tooling (LangSmith, Langfuse, Phoenix, or equivalent) Answer these in your proposal (proposals without answers are ignored): 1. Describe a multi-agent system you took from unstable to production. What was actually broken, and how did you prove it was fixed? 2. How do you prevent infinite loops and bad handoffs between agents? 3. Which orchestration framework would you pick for this, and why not the others?