Agent Reflection · Comparison
Your standard lives in every agent's repo. Which copy is current?
A critique prompt in the repo is the quickest way to start, and for one agent with one standard it may be all you need. This page compares it with a shared catalog and a separate reviewer, and says where the copy in the repo starts to cost more than it saves.
What writing your own does well.
You control every word. There is no new vendor and no new account, and the first version takes an afternoon: a prompt that asks a model to review the draft against a checklist, and a loop that runs it again. For one agent and one standard, that is a reasonable place to stay.
Where a copy in each repo stops.
The second agent gets a copy of the first agent's prompt, and the third gets a copy of the second. Each one drifts, and when the brand guide changes, someone has to find every copy.
The reviewer is often the same model, with the same instructions, that wrote the output, so it can share the writer's blind spots.
Around the prompt sits everything else you build and keep running: quoting the failing passage in a form the agent can use, a cap on attempts, versions of the standard, a log of results, and a way for the person who owns the standard to read it. None of it is hard. All of it is yours to maintain.
What Agent Reflection does instead.
Agent Reflection moves the standard out of the repo and into one catalog. Rules are Markdown files whose versions never change. The review is run by VindexAI, a reviewer separate from the agent that wrote the output, and each result is pass or fail for each rule with the failing passage quoted. The attempt cap, the log of results, and the reports come with it.
Side by side.
| What you need | A prompt in each repo | Agent Reflection |
|---|---|---|
| Setup | An afternoon for the first prompt | Connect an MCP client, write or add a rule, run a critique |
| Where the standard lives | In each agent's repo | In one catalog, read by every agent |
| Changing the standard | Edit every copy | Add a new rule version. A rule set keeps using the version it pinned |
| Who can read and edit it | Whoever can open the repo | Whoever owns the standard, in plain language |
| Who reviews the output | Often the same model that wrote it | VindexAI, separate from the agent |
| Failing passage quoted | If you build it | Yes |
| Attempt cap | If you build it | Yes |
| Reports | If you build them | Pass rates, rule fail rates, and credits per passed session, with Track |
| Cost | Model calls and engineering time | $0.03 per critique, and a new account starts with 100 free critique credits |
The middle column describes a critique prompt kept in each agent's repo. This page describes a category in general, not any one product.
When to move.
Start in the repo if you have one agent. Move to a catalog when a second agent needs the same standard, when someone outside engineering owns it, or when you need to show a pass rate. Your existing prompt becomes your first rule: write the checklist as a Markdown file, test it, and set it to active.
Questions.
- When is a critique prompt in the repo enough?
- When one agent follows one standard, the standard changes rarely, and the person who owns it works in the repo. Past that, the copies drift.
- Can I move my existing prompt into Agent Reflection?
- Yes. Write the checklist as a rule in Markdown, test it against outputs you already have, and set it to active when it behaves. The guide on writing your first rule walks through it.
- Who reviews the output?
- VindexAI, which is separate from the agent that wrote the output. Each critique is billed as one call at $0.03.