Chaser Agent: designing a harness that refuses to be trusted

Chaser Agent

Why the first version of Chaser Agent deliberately has no authority, and what a review-first output artifact has to separate to be worth reading.

Source: Public repository: chasedndt/Chaser-Agent

Context

Chaser Agent started as an extraction from ChaseOS: take the part that reads sources and reasons about them, and make it stand alone. The temptation with an agent that reads sources is to let it act on what it finds. V0 goes the other way.

Problem

An agent that reads a source and returns a confident paragraph destroys the one thing a reviewer needs: the seam between what the source said, what the agent inferred, and what the agent is unsure about. Once those are blended into prose, verification costs as much as doing the work yourself.

Requirements

The output had to keep source claims, evidence snippets, uncertainty, action candidates and memory candidates as separate, individually reviewable fields. A reviewer should be able to reject an inference without discarding the evidence underneath it.

Constraints

No provider routing, no browser control and no runtime authority by default. The harness produces artifacts; it does not take actions. Memory is proposed, never promoted automatically.

Trade-offs

Structured output is more verbose and less immediately readable than a summary. That cost is accepted deliberately: the artifact is optimised for a human deciding whether to trust it, not for a human skimming it.

Current status

Source Card Harness V0 is complete and produces deterministic local review artifacts. It is not a production autonomous agent and is not described as one.

Next milestone

Evaluation harness: a way to measure whether the separation actually improves reviewer decisions, rather than assuming it does.