# CipherClaw, debugging the agents that other tools cannot see.

> Case study: CipherClaw, open-source debugging and observability for multi-agent AI systems. Apache 2.0. Built by Alex Cinovoj.

Canonical: https://alexcinovoj.com/case-studies/cipherclaw

An open-source debugger and observability layer for multi-agent AI systems. Built when production teams kept asking the same question: why did the chain stop?

Apache 2.0. Used by teams running Claude, OpenAI, and mixed-model agent stacks.

## Multi-agent failures are invisible by default.

When five agents pass tasks between each other, a single bad tool call cascades into silence. Logs scatter across processes. Traces stop at the model boundary. Teams ship blind.

## Trace every agent, every tool, every retry.

### Span-level traces

Every agent call, tool invocation, and model response captured with parent-child links.

### Replay

Re-run a failed chain locally with the exact inputs that broke production.

### Eval harness

Pin golden tasks. Regress on every model upgrade.

### Open source

Apache 2.0. No vendor lock-in, no surprise bills.

## Production multi-agent debugging without the rebuild.

Adopted by teams shipping Claude agents in finance, legal, and customer ops. Listed in the Anthropic ecosystem. Powers internal observability across my own production AI work.

## Common questions.

## Have a multi-agent system stuck in pilot?

Hardening a multi-agent system into something your team can own runs through Systems Decision Audit at TechTide AI. CipherClaw is one of the tools I use to find the blockers. Implementation notes live on alexcinovoj.dev.