What does a critique return?

A pass or fail for each rule, with the passage that broke the rule quoted, the rule versions that ran, and your credits remaining. For a rule set you also get one pass or fail for the whole set, which passes only when every rule passes, and the list of rules that failed.

Can one call check several rules?

Yes. A rule set checks the output against every rule in it in one call, and costs the same as checking a single rule.

What happens when a rule keeps failing?

The agent revises and runs again until every rule passes, it reaches the attempt limit, or it runs out of credits. It then stops and reports which rule is still failing.

Does Vindex rewrite the output?

No. Vindex judges the output and quotes what broke a rule. Your agent does the rewriting.

Do I have to change my agent?

No new framework. Your agent connects to the Vindex MCP server like any other tool, and a short standing instruction tells it when to check.

Agent Reflection · How it works

From a written rule to a checked output.

Agent Reflection does two jobs. It keeps your rules in one place where every agent can read them, and it checks agent output against those rules before the output ships.

The loop

Read the rules. The agent calls list_rulesto see the rules it may use: your company’s own, and public rules. Reading rules is free.
Check the output. When the agent has a finished draft, it calls validate with the output and a rule or a rule set. VindexAI judges the output against each rule. A critique uses a credit.
Fix and run again. For every rule that failed, the agent rewrites the quoted passage and calls validate again. It stops when every rule passes, when it reaches the attempt limit, or when credits run out, and then tells you what is still failing.

Your rules

A rule is a written requirement an output must meet, in plain language: no exclamation marks, every claim cites a source, every acceptance criterion can be tested. Rules belong to your company, not to a person, and they are private to your company.

Versions. Editing a rule saves a new version. Earlier versions stay readable, and past results still point at the version they ran against.
Test before you rely on it. A new rule starts as a draft, moves to test while you check it by hand, and goes live when you set it active.
Where you manage them. In Agent Rules in your portal, or through your agent, which can create and update rules for you.

Rule sets

A rule set groups rules, each pinned to a version, so one call checks the output against all of them. The rules are judged in parallel, and a rule set costs the same as a single rule. Editing a rule does not change a rule set until you save a new version of the set that pins the new rule version.

What comes back

For each rule you get pass or fail. When a rule fails, the passage that broke it comes back quoted, so the agent knows exactly which words to change. You also get the rule versions that ran and the credits you have left.

For a rule set you also get one pass or fail for the whole set, and it passes only when every rule in it passes. The failed rules are named, so the agent knows which to fix first. The judgement is probabilistic, so the same output can occasionally get a different result on another run.

See a sample critique.

Who reviews the output

VindexAI, Vindex’s own model, runs every critique. No third-party model is involved. The reviewer is separate from the agent that wrote the output, so the agent does not grade its own work.

Track: see which rules your agents fail

Track opens the stats. Sample the outputs you want recorded and see pass rates by rule, by agent, and by rule version, with drift alerts when a rule starts failing more often, and exports for your own reports. Track is $49 per month, and each sampled critique is $0.03.

Languages

Rules and outputs work in 128+ languages. Write a rule once and check output in the language your agent wrote it in.

How it compares

Agent Reflection works beside the tools you may already use. Each page below says what the other category does well, where it stops, and how the two fit together.

Questions

What does a critique return?: A pass or fail for each rule, with the passage that broke the rule quoted, the rule versions that ran, and your credits remaining. For a rule set you also get one pass or fail for the whole set, which passes only when every rule passes, and the list of rules that failed.
Who reviews the output?: VindexAI, Vindex's own model. No third-party model is involved, and the agent that wrote the output does not grade it.
Can one call check several rules?: Yes. A rule set checks the output against every rule in it in one call, and costs the same as checking a single rule.
What happens when a rule keeps failing?: The agent revises and runs again until every rule passes, it reaches the attempt limit, or it runs out of credits. It then stops and reports which rule is still failing.
Does Vindex rewrite the output?: No. Vindex judges the output and quotes what broke a rule. Your agent does the rewriting.
Which languages are supported?: Rules and outputs work in 128+ languages.
Do I have to change my agent?: No new framework. Your agent connects to the Vindex MCP server like any other tool, and a short standing instruction tells it when to check.