Skip to content

Demo 03 · Recorded + simulated

Agentic trace viewer

Debugging an agent system is mostly reading traces like these — knowing which step burned the tokens and which retrieval went wrong before touching the code. Pick a scenario and replay the agent loop step by step: the tool choice and the reasoning behind it, the retrieval hits behind each observation, and the step budget counting down.

Every retrieval runs over the same 12-passage corpus as the search demo, passages about my own projects — read the case studies →

The traces marked recorded are real runs: Qwen2.5-0.5B-Instruct, a small open-weight model, running locally on CPU through an OpenAI-style tool-calling harness (recorder source). Every step is what the model actually did, unedited; latency and token counts are measured. Its answers are rough — repeated queries, misattributed sources — which is what a 0.5B model under budget pressure honestly produces, and the readable loop around it is the point. The traces marked simulated are hand-authored, with authored latency and token estimates.

Presets

A comparison question needing retrieval across projects — recorded from a real local-model run.A multi-project summary against a tight step budget — recorded from a real local-model run.A question the corpus does not directly answer — recorded from a real local-model run.Two retrievals chained — the second query is scoped by what the first one found.The step budget runs out before the last retrieval — the answer ships partial, with a caveat.The first retrieval misses; the agent reformulates and cites the nearest evidence honestly.

A comparison question needing retrieval across projects — recorded from a real local-model run.

Which of these projects handled the most data, and how was it made fast?

RECORDED · Qwen2.5-0.5B local

onnx-community/Qwen2.5-0.5B-Instruct · q4 · transformers.js (onnxruntime) CPU · sampling, temperature 0.3, top_p 0.9

budget 6/6 · 0.0s · 0 in / 0 out

Press Play to replay the trace, or Next to step through it.