Agentic AI and Infrastructure
Chaser Agent
A review-first agent harness that turns a bounded source into evidence-linked work you can inspect before anything gains authority.
Outcome
Start with one question and one source. Chaser Agent produces eight traceable artifacts—claims, evidence, uncertainty, proposals, review material and a run log—without silently executing or promoting memory.
Working proof · Chaser Agent in use
One improvement loop, two open-source tools.
This is agent evaluation and harness refinement: a human checks the work, labels the result and preserves the correction so the system can be improved and tested again. It does not change model weights.
Chaser Agent
Keep the work, not just the answer.
Chaser Agent writes an immutable review bundle: the task, source card, claims, evidence, proposed actions, memory candidates, uncertainty and run log.
Agent Review Studio
Put each judgement beside its proof.
The Studio validates the files and links, then guides the operator through every artifact group. Five quality dimensions are scored from 0 to 3, with a decision and correction notes saved as a new revision.
Engineering loop
Turn trusted corrections into repeatable tests.
Approved reviews become golden cases and regression tests. Engineers update prompts, tools, retrieval, memory rules, approval gates or orchestration, then run the same case again and compare the result.
Meet the companion
A clear signal.
At every step.
Watch ten runtime states in motion, or choose any state yourself. Sound stays off until you enable it.
From following a source to asking for your approval, the companion makes its state visible. A visual cue explains the state; it never grants permission.
Explore the interactive workspace →
Media
Select any frame to enlarge
01 · Character identity
The story begins with the finished neutral character: one consistent silhouette, surface language and visor from every angle.
- Ready
Approved Chaser Agent identity shown through a complete 360-degree studio turntable.
02 · Review-first runtime
The approved character now appears inside a concrete product story: inspect the state in 2D or 3D, follow the example workflow and pause before anything gains authority.
- Ready
Review-first runtime film — one evidence packet moves through compilation, a visible human approval boundary and a completed result with provenance intact. The real one-time run starts with one bounded source and creates eight local artifacts for inspection. Claims stay linked to evidence, human review is recorded separately and the run log proves that no external action occurred.
03 · Follow the source
Chaser moves from quiet availability to a visible investigation. The provenance trace shows what is being followed and where it is heading.
Dormant — calm and available, with no active provenance path competing for attention. Tracking — linked evidence advances toward a target that has not yet been reached.
04 · Work through the evidence
Evidence can converge, remain uncertain, or contradict itself. Each condition stays visible instead of being flattened into one confident response.
Compiling — teal and violet inputs converge in a floating working projection, kept visibly separate from the body. Uncertain — an amber question and open alternatives make the unresolved choice explicit. Contradiction — competing evidence paths stay separate and visible for review.
05 · Put the human gate in view
A recommendation can reach the decision point, but it cannot promote itself. The path changes only when the operator decides.
Awaiting approval — the provenance path stops at a visible operator checkpoint. Approved — the path continues after an attributed human decision, with the proof object still visible.
06 · Close the loop honestly
The final state says what happened: repair the route, stop at a boundary, or settle the completed result without erasing its provenance.
Correction — the wrong route retracts while evidence is reordered into a reviewable replacement. Blocked — red is reserved for a route that has stopped at a protected boundary. Complete — the result settles calmly while its supporting provenance remains inspectable.
Case study
The problem
Most agent demos optimise for speed and autonomy, which collapses source facts, model inference, recommended actions and memory updates into one confident blob of prose. Once that happens, a reviewer cannot tell what is grounded, what is speculative, what is safe to act on, or what deserves to become durable knowledge — so verification costs as much as doing the work by hand.
Behaviour before implementation
The repo is deliberately built in thinking order: a Layer 0 Behaviour Contract — the product constitution — comes first, then a V0 definition and blueprint, with a 17-layer architecture map kept subordinate to the contract. Code only exists where behaviour was defined first, and existing tests were reclassified as smoke or schema checks unless they genuinely test contract behaviour.
What V0 produces
The Source Card Harness now has a real browser-local one-time run. Paste or load one safe text source and it emits eight deterministic artifacts: a source card, claims table, traceable evidence snippets, explicit uncertainty labels, action candidates, memory candidates, a human review packet and a run log. It calls no model provider, browses nothing, promotes no memory and executes no action.
Testing and evals
Twenty-eight tests across ten suites verify the harness, the packet schemas, skill gating, research intake and the repo scaffold itself. Five golden JSONL datasets — action extraction, citation grounding, memory candidates, source-card summaries and a trading-research workflow — seed the executable checks. Agent Review Studio now adds the missing operator layer: it imports the complete run, puts claims beside their evidence, records five anchored quality scores and preserves each correction as a new review revision.
The governance boundary
Chaser Agent may propose; ChaseOS decides. Canonical truth, permission boundaries, runtime authority and promotion rules stay with the parent control plane. The harness's outputs are review-only by design, which is the point: an artifact optimised for a human deciding whether to trust it, not for skimming.
Where it is now
Phase 1 — Source Card Harness V0 — is complete and open source under MIT. The public runtime workspace can now build and export a real bounded review bundle in the browser, while Agent Review Studio remains the deeper evaluation lane. This is not production-ready autonomy, foundation-model training or a canonical truth engine; live providers, browser authority, an HTTP service and managed hosting remain future engineering lanes.
Verified architecture
System map
Technology
- Python
- React
- Deterministic review artifacts
- Evidence modelling
Scope and boundaries
The public P0.1 run is local and deterministic: no model provider, browsing, memory promotion, dispatch or external execution. It is not a production autonomous agent or browser-control adapter.