Agentic AI and Infrastructure
Agent Review Studio
A workspace for reviewing what AI agents did, checking their work against evidence, and turning trusted runs into reusable examples that help improve future agent behavior.
Outcome
Provides the review and improvement system around an agent: import complete runs, find mistakes, score the result, record corrections, compare revisions and build trusted examples and repeatable tests for prompts, tools, memory and workflows.
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.
Media
Agent review workspace
Review an agent's output and supporting evidence side by side, then score the complete run and record what should change. Dark mode keeps the same project, review and scoring workflow while saving the selected workspace and review progress in the browser.
Responsive operator access
On mobile, the review stages stack into a clear step-by-step flow without losing access to the imported files or saved progress.
Case study
Why it exists
AI agents can produce polished answers while still using weak evidence, making the wrong tool call or saving unreliable information. Agent Review Studio gives a human reviewer one place to inspect the agent's complete work, see what supported each decision and decide whether the run is good enough to reuse.
How a review works
Import a complete agent run and the Studio groups its outputs, sources, actions, memory choices and logs into one review. The reviewer checks claims beside their evidence, reviews the actions taken, gives the full run five quality scores and records a decision with any corrections. Automatic checks warn when files or links are broken, but the human makes the final judgement.
What it helps improve
Finished reviews become evidence-backed examples, often called golden cases. Engineers can use them to improve an agent's prompts, tool selection, search and retrieval, memory rules, approval checks and workflow logic. The exported reviews can also help select a future training dataset, but the Studio itself does not retrain the model.
Tested with a real agent harness
Version 1.0 was tested with three real Chaser Agent evaluation runs containing 24 files. The Studio imported every required file, kept each run in the correct review order and restored the workspace after a browser reload. The runs stayed marked unreviewed until a person scored them.
What Version 1.0 provides
The public Version 1.0 application includes named projects, dated review sessions, broad file import, light and dark modes, guided onboarding, saved drafts, run scoring, review history, re-review and portable exports. Team accounts, cross-device sync and the HTTP service that runs an agent are separate future systems.
Technology
- React
- Vite
- IndexedDB
- Agent evaluation
Scope and boundaries
Version 1.0 helps engineers improve an agent harness: the prompts, tools, memory rules, approvals and workflows around the model. It does not retrain model weights, run agents, share data between browsers or approve changes automatically.