Show HN: Agent Firewall – Go proxy to kill LLM death spirals

I run 6 AI agents as my entire team. Yesterday two agents got stuck in an infinite loop arguing over JSON formatting. Burned $47 in API calls while I slept.

Anyone running multi-agent setups (CrewAI, AutoGen, LangGraph) knows the pain: agents go rogue, tokens burn, no circuit breaker.

Building an open-source Go reverse proxy. Change one env var (OPENAI_BASE_URL=http://localhost:8080/v1), it detects death spirals and physically cuts the connection.

Local-only, no SDK, no cloud, single binary. Ship in 2 weeks if enough people need it.

Have you been burned by runaway agents?

Exactly. We're basically living through the "distributed systems reliability" moment for agents. Traditional infra has circuit breakers, rate limiters, health checks — agents have nothing.

That's what I'm building: treat agent-to-LLM calls like microservice calls. Detect loops, enforce budgets, fail-open if the firewall itself crashes. Same patterns, new domain.

The interesting part is loop detection needs to be semantic, not just frequency-based. Two agents can politely argue in circles without ever triggering a rate limit.

[flagged]