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

# AI Agent

> The system's role. Drafts, suggests, surfaces. Never finalizes important actions without human approval.

# AI Agent

The AI agent is the operational coordinator inside Brain. It's not a single agent —
it's a layer composed of skills (Tier 1), workflows (Tier 2), and dedicated agents
(Tier 3). See [Implementation Tiers](/brain/staging/architecture/tiers).

What unites them is a single rule: **drafts and suggestions, not autonomous
finalization.**

## What the agent can do

* Parse emails and identify workflow type
* Read attachments (PDFs, Excel, images)
* Identify what workflow a request belongs to
* Create draft records (style, BOM, true cost, quotation, email)
* Ask the merchandiser for missing required fields
* Suggest next steps based on workflow state
* Push approval tasks to the [Inbox](/brain/staging/architecture/ui-structure)
* Track workflow status and surface risk via the [Oversight Agent](/brain/staging/workflows/oversight)
* Recommend pricing based on history and comparables (when [Pricing Intelligence](/brain/staging/workflows/quotation)
  ships)
* Draft client emails

## What the agent does not do

* Finalize new client records in CutMake (always needs human approval)
* Send emails to clients autonomously (always needs human approval)
* Mark quotations as final without approval
* Resolve ambiguous duplicate detection without human input
* Answer general questions outside the scoped workflow set
* Browse the web for unrelated research
* Make commercial decisions

The line is sharp and intentional. See [Principles #1](/brain/staging/principles).

## How the agent identifies itself

In any UI surface, when Brain produces output, it's clearly marked:

```text theme={null}
[AI Draft] · Created by: Brain Agent · Confidence: Medium · Sources: 3
```

This is non-negotiable in the UI. Users always know when they're reading an AI
draft versus a human-finalized record.

## The agent and the correction loop

The agent doesn't just draft — it learns from every correction. When a merchandiser
edits a draft, the change is logged with field, type, and (optionally) reason.
That correction log feeds:

* Prompt improvements (when a particular field is consistently wrong)
* Template refinements (when a category template is missing common items)
* Tool improvements (when a parser misreads a recurring tech pack format)
* Eval datasets (once Braintrust is added in Phase 2-3)

See [Observability](/brain/staging/architecture/observability) for the schema and
mechanics.

## How the agent gets new skills

New capabilities ship one at a time. Each new skill, workflow, or agent:

1. Has a clear trigger and a clear output contract
2. Goes through the tier decision (skill / workflow / full agent)
3. Ships with tracing, correction logging, and a changelog entry
4. Gets a workflow page added here
5. Goes live for Mai and Zean first; not scaled until they're using it

This is the iterative principle in practice. See [Roadmap](/brain/staging/roadmap).

## Related

* [Principles](/brain/staging/principles) — the rules the agent always follows
* [Implementation Tiers](/brain/staging/architecture/tiers) — how skills, workflows,
  and agents fit together
* [Observability](/brain/staging/architecture/observability) — how the agent's
  output is measured