> ## Documentation Index
> Fetch the complete documentation index at: https://docs.delphina.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Critic Agent (Beta)

> Automatic review of every agent response for missing knowledge and assumptions

Every agent response is automatically reviewed by a background critic. The critic checks whether the agent's answer is grounded in your [knowledge base](/knowledge-agent/maintenance) or whether it had to guess — and flags any issues as inline annotations on the response.

<Note>
  The critic agent is in beta. Its checks and annotation format may evolve.
</Note>

## What it checks

* **Missing knowledge** — the agent answered using schema inference rather than documented definitions.
* **Unjustified assumptions** — the agent guessed a calculation without a documented basis.
* **Ambiguous question** — your question could be interpreted multiple ways.
* **Incorrect reasoning** — a logical issue with the query or interpretation.

If the critic finds nothing to flag, no annotations appear.

## How annotations look

Annotations appear inline on the response within a few seconds:

> **You:** What was our MRR last month?
>
> **Agent:** MRR for February 2026 was \$1.8M, up 5% from January.
>
> **Critic — Missing knowledge:** *The agent used `payments.amount` to calculate MRR. No MRR metric is documented. Consider documenting MRR with `/knowledge` so the agent uses a consistent definition.*

## What to do

| Annotation type             | Action                                                               |
| --------------------------- | -------------------------------------------------------------------- |
| **Missing knowledge**       | Document the metric or table with `/knowledge`                       |
| **Unjustified assumptions** | Add the business rule or definition with `/knowledge`                |
| **Ambiguous question**      | Rephrase with more specifics (time range, metric name, filters)      |
| **Incorrect reasoning**     | Review the SQL, then update the knowledge base if the logic is wrong |

## Relationship to issues

Critic annotations automatically generate issues that appear in the **Context Layer > Inbox**. This means you don't have to act on every annotation immediately — you can triage them later alongside eval failures in one place. See [Evaluations & Quality](/trust-and-quality/evaluations) for the full workflow.