# Controlled Browser Agent Stack

> A controlled technical stack for browser-only tasks where APIs are unavailable, with scoped access, replayable evidence, and approval gates.

- Canonical: https://gptnavi.com/stacks/controlled-browser-agent-stack
- Estimated monthly cost: $0-$500/month plus browser and model usage
- Last materially updated: 2026-08-24
- Best for: Operations teams, AI engineers, Technical founders

## Problems this stack solves

- Browser-only work
- Brittle selectors
- Unclear agent actions
- Credential exposure
- Irreversible submissions

## Recommended tools

### Agent framework

Recommended: Browser Use, Stagehand

Why: Combine flexible natural-language execution with explicit, inspectable browser steps.

### Browser environment

Recommended: Steel, Browserbase

Why: Use isolated sessions, profiles, observability, and replays rather than a broad personal browser.

### Controlled tools

Recommended: FastMCP, PydanticAI

Why: Expose only the actions and fields an agent needs, with typed validation.

### Human approval

Recommended: Slack, Linear

Why: Pause submissions and sensitive changes until a named person approves the evidence.

## Beginner setup plan

1. Prototype against a sandbox or read-only task first.
2. Use separate, scoped credentials for the agent.
3. Alert on unexpected navigation or action attempts.
4. Add write actions only after several reviewed read-only runs.

## Included workflows

- https://gptnavi.com/workflows/browser-agent-with-approval-gates
- https://gptnavi.com/workflows/browser-task-to-sop-and-automation
- https://gptnavi.com/workflows/tender-and-vendor-page-monitoring