ARGUS
agent observability lab
Ask a shop agent.See every step.
Argus runs a tool-calling agent against a deterministic replay of a fictional field-supply shop. The answer is reproducible. The OpenTelemetry trace is real.
The loop
- 01
Prompt
The browser sends one message and the conversation history.
- 02
Agent
A LangChain agent chooses a tool and records a GenAI span.
- 03
Tool
The tool calls the local shop API over HTTP.
- 04
Trace
Jaeger shows the LLM call, tool span, and token counts.
replay default
No provider key is needed for the public demo.
server-side BFF
The agent URL never reaches the browser.
Pick a recorded prompt
Replay answers these three prompts. Every reply exposes the tool call, the trace ID, and the token estimate behind it.
Replay answers the three recorded prompts. Use New chat between prompts; the agent URL stays server-side.