Agent Reflection
Agent Reflection keeps the rules your AI agents must follow and checks their output against them. This page defines the words used across the docs and points to where each tool is documented. To run your first critique, start at Connect your agent.
The words
| Term | Meaning |
|---|---|
| Rule | A written requirement an output must meet, saved as a Markdown file with a name, a description, and tags. Rules belong to your company, never to a person. |
| Rule version | Each complete upload of a rule. A version never changes after it is saved, and earlier versions stay readable. |
| Status | Where a rule is in its life: draft, test, active, disabled, or archived. Only active rules are listed for your agents by default. |
| Catalog | Your company's collection of rules and rule sets. One per company, private by default. |
| Rule set | A named, versioned group of rules, each pinned to a rule version. One call checks an output against every rule in it. |
| Public catalog | Rules and rule sets published for any company to use. Vindex publishes reference rules there. |
| Critique | One output checked against one rule or one rule set. The MCP tool that runs it is validate. |
| Critique credit | The billing unit. A critique costs $0.03 and uses one credit. A new account includes 100 free credits, once. |
| Session and attempt | A loop of checks on one rule or rule set. Each call in the loop is an attempt, and the loop stops at a cap. |
| VindexAI | Vindex's own model, which runs every critique. No third-party model is involved. |
Ways in
- Your agent, over MCP. One server, one sign-in. The same server also scores user stories for Integrated Vindex. The MCP server reference covers the endpoint and sign-in.
- Agent Rules in your portal. Upload a rule, tag it, change its status, and manage many at once. A rule made in the portal is visible to your agents straight away, and the other way round.
The tools
Thirteen tools in sixfeatures. Access is set per feature, by role. Items another company owns and has not made public return “not found”, never “forbidden”.
| Feature | Tools | What it does | Details |
|---|---|---|---|
| Account | get_account, get_credit_balance | Who you are working for, your plan, and your credit balance | MCP tools |
| Rules | list_rules, create_rule, update_rule | Find, write, and version rules | Rule format |
| Rule sets | list_rulesets, create_ruleset, update_ruleset | Group rules and check an output against all of them in one call | Rule sets |
| Governance | approve_item | Company admins only: control which items your agents may use | MCP tools |
| Validation | validate | Check an output against a rule or a rule set | Critique results |
| Reporting | get_reporting_schema, query_report, get_benchmarks | The stats on how your agents do against your rules (Track) | Reporting |
Languages
Rules and outputs work in 128+ languages. Write a rule once and check output in the language your agent wrote it in.
What Vindex does not do
- It does not rewrite the output. It judges the output and quotes the passage that broke a rule. Your agent does the rewriting.
- It does not block output. A critique returns a result, and your agent or your workflow decides what happens next.
- It is not a safety layer. It is not built to catch personal data, injected instructions, or format errors.
- It does not certify anything. A pass means the output met the rule as written.
- It does not review code. Rules check text.