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

# Your agent's context

> The documents your agent runs on, how to teach it a new skill, how to protect a change with a test, and how a change goes live.

Your agent's behavior comes from a set of documents we call its **context**:
who it is, how it speaks, its rules, the situations it knows how to handle, and
the facts about your business. You change them in **Agent Studio**, or from an AI
tool connected over [MCP](https://docs.useboom.ai/use-mcp.md). Either way, a change goes to a **draft**
first and reaches your customers only after it is reviewed and approved.

> [!NOTE]
>
> Editing the context from an AI tool is being enabled per organization. If the
> tools answer that it isn't available for you yet, ask your Boom contact.

## What the context is made of

| Area | What it governs | Example change |
| - | - | - |
| **Persona** | Who the agent is: its name, role, what it represents | "Introduce yourself as Sofi, from Aromante's team" |
| **Voice and tone** | How it writes: formality, length, emojis | "Shorter replies, no exclamation marks" |
| **Policies** | What it can promise, must refuse, or has to hand to a person | "Never promise a delivery date" |
| **Glossary** | Your terms and what they mean | "MSI means meses sin intereses" |
| **Knowledge** | Facts about your business: products, prices, hours, shipping | "Free shipping from \$1,198" |
| **Procedures** | How it handles a situation, start to finish | "When a customer asks to return an order..." |
| **Channels** | Differences per channel, such as WhatsApp vs email | "On email, sign with the team name" |
| **Skills** | Step-by-step tasks the agent performs, such as looking up an order | Adjust a step of the order lookup |

The areas are fixed: you change and add to what is inside them, but you can't
create a new area.

### When the agent reads each document

Not every document is read on every message:

* **Every turn**: usually persona, voice and tone, policies and the core
  knowledge. A rule that must always hold belongs in one of these.
* **Only when it applies**: each procedure is read when a customer's message
  matches its trigger ("when a customer asks for a refund"). This keeps the
  agent focused, but a rule written inside a procedure only applies in that
  situation.

Agent Studio and the MCP tools show which documents are read every turn and
which only when their procedure fires.

## Teach your agent a new skill

To make the agent handle something new, add a **procedure**: a situation, what
triggers it, and the steps to follow. Procedures are how you give your agent a
new skill.

1. **Describe the situation** as something the customer does: "a customer wants
   to return an order", not "process returns".
2. **Write the trigger** in your customers' own words, including what it is
   *not* for if another procedure could be confused with it. A vague trigger
   never fires.
3. **Write the steps** as actions the agent takes ("ask for the order number",
   "if it was more than 30 days ago, offer store credit"), not goals.
4. **Add a test** (below), then send it for review. For now, tests are added
   from an AI tool connected over MCP.

Ask Agent Studio to add it ("when a customer asks to return an order, ..."), or
an AI tool connected over MCP (`org_context_add_procedure`).

## Protect a change with a test

A **test** is a short, invented conversation plus the points the agent's reply
must meet, such as "offers store credit" or "does NOT promise a refund". When a
change is sent for review, your agent's tests run against it and the reviewer
sees the results before approving. That's how a change that would break an
earlier rule gets caught.

* Add a test for **every new procedure and every rule** you add or change.
* **Invent the conversation.** Never paste a real one or use real names, phone
  numbers or emails.
* Test what the reply must do, not its exact words. One situation per test.

Tests are added from an AI tool, with `org_context_add_eval`; adding them from
Agent Studio is coming. Reviewers expect a test with every new procedure.

## From draft to live

1. **Change**: each edit is saved to your organization's draft. Nothing your
   customers see changes yet.
2. **Send for review**: in Agent Studio, open the unpublished-changes indicator
   and choose **Send for review**; from an AI tool, `org_context_publish`. Your
   tests run against the draft.
3. **Approve**: a reviewer approves the change, and your agent is rebuilt with
   it. Agent Studio shows when it's live.

You can look at or throw away an unsent draft at any time.

## From an AI tool (MCP)

Connected over [MCP](https://docs.useboom.ai/use-mcp.md), an AI tool such as Claude can do all of the above:

| Tool | Does |
| - | - |
| `org_context_guide` | Returns the guide to changing the agent. Every other change tool requires reading it first. |
| `org_context_list`, `org_context_read` | List the documents and read one |
| `org_context_edit`, `org_context_edit_batch`, `org_context_setting_set` | Change text or a setting, saved to the draft |
| `org_context_add_procedure` | Add a new skill (procedure) |
| `org_context_add_eval` | Add a test |
| `org_context_draft_get`, `org_context_publish`, `org_context_discard` | See the draft and how it did on your automatic checks, send it for review, or throw it away |
| `conversations_get`, `agent_session_get` | Read a conversation and what the agent did on each turn (including which of your documents it opened), before deciding what to change |

The same rules apply as in Agent Studio. Changing the context takes the
**Owner** or **Admin** role, and so does reading conversations and the agent's
trace from an AI tool. The [Boom skills](https://docs.useboom.ai/skills.md) for Claude include
`edit-your-agent`, which teaches it this whole process.

## Related

> [card] [The Agent](https://docs.useboom.ai/the-agent.md)
>
> What answers your customers, and how it improves from your corrections.

> [card] [Use Boom from an AI tool](https://docs.useboom.ai/use-mcp.md)
>
> Connect Claude or another AI tool to your organization.


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