A LangGraph agent fails on step 8. The trace shows exactly what broke. To reproduce it, you have to run the whole thing again: eight more calls to the model, about 30 seconds, and another 15 cents in API costs. If the failure came from the model giving an unusual response, you may not be able to reproduce it at all. You end up debugging something that will not hold still.

We built Agent Observability to fix that. It is a Python library and command-line tool that records every network request your agent makes to a small local file, then replays those exact bytes offline, in about 0.93 milliseconds on average, compared to roughly 8,500 milliseconds for a real run against GPT-4o across 10 steps. Recording adds about 0.011% overhead, roughly 0.09 milliseconds per call to the model. The replayed result is identical, byte for byte, to what was recorded. You record once, and you can replay as many times as you want, for free.

Record a run with @tracer.instrument(record=True) around the call you want captured. Replay it with the replay() context manager, same code path, zero network, identical result. In CI, set AGENT_TRACE_NETWORK_GUARD=1 so any test that tries to make a real network call fails loudly right away, instead of quietly using up your API budget.

Agent Observability is not meant to replace dashboards or evaluation tools. It solves one specific, earlier problem: reproducing a failed run without spending any more money on the model API, for any agent built on any Python HTTP client.

Repo: https://github.com/RudrenduPaul/agent-observability
MCP Servers:
https://mcpservers.org/servers/rudrendupaul/agent-observability
https://glama.ai/mcp/servers/RudrenduPaul/agent-observability
NPM: https://www.npmjs.com/package/agent-observability-trace-cli
PyPI: http://pypi.org/project/agent-observability-trace-cli

We would love to hear your feedback and answer any technical questions on this topic. Let us know what you think!