# Governed agent with connected tools

> Build a narrow internal agent that can retrieve approved context, propose actions, and use connected tools only after explicit policy and approval checks.

- Canonical: https://gptnavi.com/workflows/governed-agent-with-connected-tools
- Category: Operations
- Difficulty: Advanced
- Setup time: 5 hours
- Estimated time saved: 4-12 hours
- Last materially updated: 2026-08-19
- Best for: Operations teams, AI engineers, Customer success teams
- Tools: LangGraph, LlamaIndex, Composio, Smithery, PydanticAI

## Quick answer

Connected agents create real leverage and real operational risk. Start with read-only evidence and draft actions, then add a small set of approved actions with a named owner.

## When to use it

- Internal policy assistant
- Account research draft
- Support triage
- Operations checklists

## Steps

1. **Define the agent boundary** — Specify one user request type, approved sources, disallowed content, allowed actions, required approval, and the escalation owner. Tool: PydanticAI. Expected output: An agent policy contract.
2. **Connect only approved knowledge** — Index a small, permissioned source set and require answers to return citations or clearly say that evidence is missing. Tool: LlamaIndex. Expected output: A cited knowledge layer.
3. **Model state and approval** — Build a workflow that separates research, draft action, approval, execution, and audit record. Tool: LangGraph. Expected output: A reviewable agent graph.
4. **Add selected business actions** — Connect only the minimal tools needed, with scoped credentials and confirmation before any external impact. Tool: Composio. Expected output: A controlled action surface.
5. **Review tool contracts** — Use well-maintained tool servers and test each integration against policy, permissions, error handling, and logging requirements. Tool: Smithery. Expected output: An approved integration checklist.

## Prompt templates

### Agent policy

Write an operating policy for an internal AI agent. Include supported requests, allowed sources, required citations, prohibited content, permitted actions, confirmation rules, escalation triggers, audit fields, and named owner. Use case: [paste]

### Action safety review

Review this proposed agent action. Identify missing evidence, permission issues, irreversible impact, customer risk, data exposure, and whether human approval is required. Proposed action and context: [paste]

## Common mistakes

- Giving write access before citation quality is proven
- Using broad credentials for convenient integrations
- Treating a tool call as proof an action was correct

## Related workflows

- https://gptnavi.com/workflows/governed-knowledge-agent-for-account-and-policy-questions
- https://gptnavi.com/workflows/ai-agent-ops-from-recurring-task
- https://gptnavi.com/workflows/llm-evaluation-before-production