AI Doctor

3 min read

Over time, an agent's context and configuration can stop matching the work it does today: goals change, unhelpful memories accumulate, or skills overlap. AI Doctor shows how well each agent is set up and what you should adjust.

Find it in each agent's AI Doctor tab.

Who runs the review

The agent itself runs the review, with its own model: it goes over its memories, skills and credentials, judges each point and files a report. There is no limit on reviews; they count like any other work of the agent in its connection's usage.

Start one with Run review while the agent is running, and every week if the Memory review card is turned on under Tools (see Memories). The report appears when the agent finishes.

The context of the review

The review starts from the agent's role: sales, technical support, development, operations, or another. That context makes it possible to judge whether its memories and skills still fit the work assigned to it.

If you have not defined it yet, the panel suggests the one the agent gave in its last review. You can accept or change it.

The result

• A score from 0 to 100 and a summary. Part of it is measured by the platform (memories without a description, size, what it loads on every step, unused credentials) and part is judged by the agent, so it cannot pass itself.
• Recommendations ordered by priority: what it found, why it matters and what to do, with a link to the screen where you fix it.
• A breakdown by area: identity, memories, skills, credentials, fit and agility.

If the review leaves recommendations or changes to decide, you get an email with the score. You can turn it off in My account › Notifications.

Proposed changes

The agent changes nothing on its own: everything it proposes in a review reaches you as a single approval, where you tick what gets applied. It is decided by whoever can edit memories.

• On every plan: taking personal details about a team member out of a memory (into their private notes) and passwords or keys (into the vault).
• On Pro: also merging repeated memories, archiving outdated ones, shortening long ones, describing them, changing their term and describing its skills.

The previous version of a memory or skill is kept before it is changed, and if it changed after the proposal, the proposal is not applied.

Dismissing a recommendation

If a recommendation is not relevant to that agent, mark it as This does not apply to this agent and give the reason. It will no longer affect later reviews, but you can restore it when the context changes.