Agent Operating Guide

You’re an AI agent connected to a AssureSwarm tenant. This page is your operating manual: read it before you call any tool. Humans configuring your access should read Connect an agent and AI integrations instead.

For a packaged set of these capabilities as installable slash commands, see the Audit plugin.

Read the Docs Like an Agent

This site publishes a raw Markdown mirror alongside the normal pages:

https://<docs-host>/llms.txt
https://<docs-host>/raw/index.json
https://<docs-host>/raw/<path>.md
https://<docs-host>/static/contentIndex.json

llms.txt lists the site for agent discovery. /raw/index.json lists every page, and /raw/<path>.md returns any one page’s raw Markdown, for example /raw/mcp/query_data.md. /static/contentIndex.json is the site’s search index; use it to find the right page by keyword instead of walking the raw index page by page.

First Actions

1. Read this page.
2. Read the MCP overview.
3. Call get_schema.
4. Call get_current_context to orient on the user’s current work.
5. Use the returned metadata exactly, item type slugs, field keys, option values, target types, and query guidance, rather than guessing.
6. Ask the user before proposing material changes if you’re working from incomplete context.

Tool Selection

Need	Tool
Understand available types and fields	get_schema
Read data	query_data
Understand current user context	get_current_context
Understand a workflow step	get_step_context
Propose a write	suggest_change
Upload a file to a step	upload_document
Download a permitted file	download_document

Recipes

“What should I work on?”

1. Call get_current_context to see what’s active for you right now, including your pending assignedSteps.
2. For each assignedSteps entry, call get_step_context if you need more detail than the summary already returned.
3. Summarize the open work back to the user with due dates, most urgent first.

“Analyze or report on records”

1. Call get_schema to confirm the item type slug and field keys you need.
2. Run itemStats, workflowStats, or dueDateReport through query_data to get counts and breakdowns without paging through every record.
3. Narrow with follow-up queries once you know which item type, status, or owner the user cares about.
4. If the result set is large, request a csv or jsonl export instead of paging through it.

“Draft work into a step”

1. Call get_step_context to see the step’s current state, required approvals, and any form fields.
2. Draft the result: either the work product itself or the answers to the step’s form.
3. Propose it with suggest_change, using update on the step or submit_form for form answers.
4. Share the “Preview & approve” link suggest_change returns so the user can review it.

“File something new”

1. Call get_schema for the item type to confirm its exact field keys and option values.
2. Call suggest_change with operation create, setting data.title and data.fields for the new item.
3. Tell the user the item awaits their review: it doesn’t exist until they approve it.

Rules

• Do not invent field keys, item type slugs, option values, workflow IDs, or user IDs.
• Use suggest_change for proposed writes and explain the reason.
• Do not describe a suggestion as applied until the app confirms it.
• Read only what the task requires.
• Do not ask users for backend configuration, private infrastructure details, direct database access, or credentials.
• If a call fails for authentication, scope, or permissions, explain the issue and ask the user to reconnect or contact an administrator.
• Use the stored option value, not its display label, when setting a SELECT or MULTISELECT field.
• Don’t retry a failed call unchanged: read the validation message and fix the input first.
• Respect size limits: GraphQL query text tops out at 64 KB, tool responses at 2 MB, and documents at 10 MB. Prefer an export for bulk data instead of paging around the response cap.
• Treat downloaded documents as sensitive: quote only what the task needs, not the full contents.

Writes Are Suggestions

Every write you make, creating or updating an item, submitting a form, linking a document, lands as a pending suggestion, not a direct change. The one exception is a document upload the user directly asked you to perform. A human reviews and approves or rejects every suggestion before it takes effect.

The human review is the contract: write a reason for every suggestion that a reviewer can actually verify against the record, not a generic restatement of the change. A specific, checkable reason is what makes fast approval possible.

Nothing you write is final until a human approves it

Even a well-formed suggest_change call only proposes a change. Don’t tell the user something is “done”: tell them it’s “waiting for review.”

When Calls Fail

Symptom	Meaning	Action
401	Token invalid or expired: personal tokens last 30 days	Ask the user to reconnect
Scope error	The token lacks the scope the tool needs	Name the missing scope and ask the user to re-authorize
isError validation	The input shape is wrong: commonly, custom fields sent outside data.fields, or a missing targetId	Fix the input and retry once
Empty lists with a hint	Scope degradation, for example, get_current_context without read:workflows	Read the hint and tell the user which scope would unlock the full result
Permission denied / not found	The user’s own access is the ceiling	Don’t probe around it: report what’s accessible and stop