The idea for Agent-Blackbox came from a real 3 AM Slack message I never want to see again. Our multi-agent pipeline had just wiped the staging database. PM-agent said the requirement was clean. Coder-agent said it followed the spec perfectly. Verifier-agent said it never even received output. Four hours of grepping through thousands of log lines later, we still had no idea which agent actually failed. That's the "accountability vacuum" — and I built Agent-Blackbox to solve it. **What problem does it solve?** When a single LLM call fails, you debug it. When a chain of 5 agents fails, you play detective. Every agent points fingers at the next one. Logs are scattered across systems. There's no verifiable chain of custody for decisions. Agent-Blackbox gives you `git blame` for AI agents. In 3 seconds, it tells you exactly which agent broke the chain — with cryptographic proof, not guesswork. **How it works (the 10-second version)** Under the hood, it implements two IETF drafts: - **JEP (Judgment Event Protocol)** — a minimal, cryptographically signed log format for agent decisions. - **JAC (Judgment Accountability Chain)** — a `task_based_on` field that links every decision to its parent. Four verbs — `J` (Judge), `D` (Delegate), `T` (Terminate), `V` (Verify) — model any accountability workflow. Every action produces a signed JEP receipt with Ed25519 signatures. When something breaks, you trace the `task_based_on` chain back to the failure point. **How it evolved** The first version was a messy internal script that just hashed logs. Then I realized the problem was bigger than my team — *everyone* deploying multi-agent systems was hitting the same wall. So I rebuilt it around IETF drafts to make it an open standard, not another closed ecosystem. v1.0 ships with: - ✅ Rust core engine (fast) - ✅ Python SDK (TypeScript coming) - ✅ `blame-finder dashboard` for visual causality trees - 🚧 LangChain / CrewAI native adapters (in progress) **What's next?** I want to make Agent-Blackbox the default accountability layer for agentic workflows — like what `git blame` did for code. PDF/HTML blame reports, real-time alerting, and deeper framework integrations are on the roadmap. **Try it yourself** ```bash pip install agent-blame-finder blame-finder dashboard ``` GitHub: https://github.com/hjs-spec/Agen... I'd love feedback — especially from anyone else who's been woken up at 3 AM by a broken agent pipeline. What's your current debugging workflow? How do you trace failures across agents today?

Forum Threads

p/agent-blackbox

•

7mo ago

Agent-Blackbox: git blame for AI agents – find who broke prod in 3s

I built Agent-Blackbox after losing a week debugging an 8-agent pipeline failure. Every agent pointed fingers at the next one, and logs were useless for tracing the actual decision chain.

What it does:

- Traces failures to the exact agent and decision point

- Reconstructs the full task-based-on chain across agents

View all

Agent-Blackbox

git blame for AI agents — find who broke prod in 3s

git blame for AI agents — find who broke prod in 3s

Forum Threads

Agent-Blackbox: git blame for AI agents – find who broke prod in 3s

Forum Threads

Agent-Blackbox: git blame for AI agents – find who broke prod in 3s