Point an agent at Jewzer and Jewzer becomes its runtime. Every answer, every tool call and every turn it decided to skip lands in an append-only log with the reasoning, the confidence and the latency attached. The agent works inside the policy you wrote. It has no way to edit the record.
| Agent | Model | 5m | Decisions | Median | Tool calls | Blocked | Status |
|---|
Before the model runs, Jewzer snapshots what the agent can see: the prompt, the retrieved context, the tools the policy allows this turn and the budget left. That snapshot is written first, so a decision can be replayed even if the run fails.
The model picks one of four moves and states its confidence: act, call a tool, wait for more signal, or halt. Anything under your review threshold goes to a human instead. Median time to an answer is 219 ms.
The move, the reasoning, the tool response and the latency are appended to the log and hashed against the turn before it. Records cannot be edited or removed, by your team or by ours. Export the chain whenever you want it.