Simulates an agentic AI session, showing where the inference engine and harness spend their time.
Useful to understand how different performance parameters affect the user experience.
A session is a sequence of turns, each turn being split in the following phases:
- queue — the request waits on the engine to start processing it.
- prefill — the engine reads the prompt and any prior context, and prepares to generate an answer.
- decode — the engine writes the answer, and streams it back one token at a time.
- tool exec — the harness runs a command and waits for its completion.
How to use
- ▶ run starts the session; ❚❚ pause pauses the session.
- The header bar displays a timeline of the whole run, with the current phase
highlighted.
Click or drag it to skip to any point in the run.
- The phase bar under the header displays the current phase and fills as that phase
runs.
Click it to step to the next phase, shift-click to step to the one
before.
- Play with the settings on the right to see how they affect the session metrics and
user experience.
Pick a pre-recorded session to start:
Inference engine options
- prefill speed — rate at which the engine can read the prompt and context.
- decode speed — rate at which the engine can generate and stream the answer.
- engine presets — a named machine as prefill, decode and request latency.
- prompt cache — whether the engine re-reads the prompt and context every turn, or
reuses them.
- request latency — how long the request waits before the engine starts processing it.
- throughput decay curve — how the engine's throughput decreases as the context grows.
Shortcuts
- space
- play / pause
- R
- restart from zero
- ← →
- seek one second, shift five
- 1 – 7
- playback speed, 0.25× – 16×
- esc
- close the panel in front