> ## 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 trustworthiness review of every agent response

After the agent finishes a response, a background **critic agent** reviews the full analysis — your question, the SQL, the query results, and the answer — and scores how trustworthy it is. The verdict appears as a **Critic** row beneath the response.

<Note>
  The critic is in beta and rolling out gradually. The review may not be visible in your workspace yet, some users see an earlier, more compact indicator during the rollout, and the checks and appearance may evolve.
</Note>

## What it checks

The critic scores four pillars:

| Pillar | Question it answers |
| - | - |
| **Clear question** | Did unresolved ambiguity in your question damage the answer you received? |
| **Grounded context** | Did the answer's definitions and rules come from sound sources the agent actually had? |
| **Sound logic** | Given the inputs it had, did the agent compute and interpret correctly? |
| **Safe assumptions** | Is the answer grounded in verified data, or does it rest on unverified choices? |

It also verifies that the numbers in the answer trace back to actual query results — a figure that can't be traced always marks the response as a likely problem.

## Reading the verdict

The collapsed Critic row is the zero-effort signal:

* **Green — "all checks passed"**: nothing flagged.
* **Amber**: something is worth double-checking; the headline of the most important finding is shown inline.
* **Red**: high probability the core answer is wrong.

Click the row to expand it. Each pillar shows its findings — a short headline with the detail behind it — and clean pillars show a one-line summary of what was verified.

The critic reviews each analysis to assess its trustworthiness; it can be wrong. Treat the verdict as a prioritization signal — a reason to look closer, not a final judgment.

## What to do with findings

| Finding under | Typical action |
| - | - |
| **Clear question** | Rephrase with more specifics — time range, metric name, filters |
| **Grounded context** | Document the missing metric, table, or rule with `/knowledge` |
| **Sound logic** | Ask the agent to correct the analysis — a logic flaw is usually the agent's error; if it traces back to documentation, fix it with `/knowledge` |
| **Safe assumptions** | Confirm or correct the assumption, and document the right definition so the agent doesn't have to guess |

Findings that trace back to missing documentation also flag the turn as having a **knowledge gap**. You don't have to act on each one in the moment — Delphina can find these flagged turns later, most naturally through a scheduled workflow that reviews them in bulk and turns recurring gaps into documentation. See [Automating maintenance with workflows](/knowledge-agent/maintenance#automating-maintenance-with-workflows).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.