Review agent
How the Clearly AI review agent works, and how to structure your sources and review templates to get the most out of it.
The review agent is how Clearly AI generates a review. It reads the sources on a project and your Knowledge Base files, works through the questions in a review template, and returns structured answers with a confidence level and any information gaps it found.
The agent explores your material the way a human reviewer would: it opens files, follows references, and searches for evidence before answering, rather than answering from a single pass over truncated text. This lets it handle large projects, such as full repositories, long design documents, and prior reviews.
How a review runs
A run moves through the stages below. You do not set these stages up. They run automatically after you click New Review, working from the review template and sources you provide.
Explore
The agent reads your project sources, your Knowledge Base files, and the review template, then maps out the system before answering anything. It notes where the evidence for each area lives and flags questions your sources may not answer.
Answer background questions
Questions your review template marks as background are answered first, one after another in the same agent conversation. Because they share a conversation, each background answer builds on the ones before it and establishes foundational facts such as the system's architecture, the data it handles, and its trust boundaries.
Order matters here: because background questions are answered in sequence, a later one can rely on an earlier one. This is the single biggest lever on review quality. Mark your context-setting questions as background, and order them from general to specific.
Fork the conversation, then answer each question
That conversation now holds the exploration plus every background answer. The agent forks it once for each remaining question. Each question is answered in its own fork, in parallel, so it inherits the full shared context but runs independently.
This is what keeps a review both consistent and fast: every answer starts from the same foundation, while many questions run at once. For each question the agent returns the reviewer-facing answer and a confidence level.
You can leave the page while a review generates and come back to it. The review screen shows live progress, then the completed answers once the run finishes.
How the agent is sandboxed
The review agent runs as a set of skill files: structured instructions that drive the explore-and-answer workflow. They run inside a network-isolated sandbox that Clearly AI creates for the run and destroys when it finishes.
The sandbox is deliberately confined:
- No general network access. It can reach only the endpoints a review needs: the model provider and Clearly AI's own services. It cannot browse the web or reach arbitrary hosts.
- Limited write path. The agent can send results back to Clearly AI only through a small, fixed set of tools: it submits an answer or reports missing information. It has no other way to act on your data or systems.
- Read-only access to sources. The agent reads the project sources and Knowledge Base files it loads, but cannot modify your repositories or other connected systems.
- Short-lived, narrowly scoped credentials. Each run gets a scoped session limited to the one review it is generating, and Clearly AI revokes that access when the run ends.
What the agent produces
Each answer includes:
- Answer: the reviewer-facing response, written from your sources. Answers can include tables and Mermaid diagrams when the question calls for them.
- Confidence: one of High, Medium, Low, or Not enough information. A low or missing-information result tells you where the sources were thin, rather than filling the gap with a guess.
- Compliance status: for questions configured with assertions, a status such as Compliant, Partially compliant, or Noncompliant. See Reviews for how status is displayed.
When the sources do not support a confident answer, the agent does not invent a response. It reports the gap as an information request on the review page, noting what is missing and why it is needed.
The agent works only from your material: it answers from project sources and Knowledge Base context, does not browse the web or pull in outside facts, and marks information as missing rather than guessing. For the broader set of controls that keep review output structured and repeatable, see AI output controls.
Regeneration
When a project's sources change after a review was generated, Clearly AI can re-review it. The agent re-examines the review against the updated sources, updates the answers that changed, and reports when nothing changed so you can see the review is still current. You can also regenerate a single answer from the review screen.