TOBIASZANDER.DE / AGENT GATESENDE

Approval gates for AI agents

Your coding agent either asks for permission every five minutes, or it can do far more than you think. Usually both, depending on the day you last touched its settings.

After the twentieth prompt you click yes without reading. So you open it up: Bash(*), bypass mode, a GitHub token with write access in an MCP server. From then on nobody checks anything, and the agent reads every issue, web page and calendar invite you point it at.

I run my own agents on my network, my VMs and my cloud accounts, and they never hold the keys. Every change goes through a gate outside the agent: harmless ones pass on their own, a second model reviews the middle ground, and only a short list of things reaches my phone, bound to the exact plan I approve. This playbook is how to build that for your own setup.

WHAT IS IN THE PLAYBOOK

1. Risk tiers instead of yes/no

Four levels (automatic, reviewed by a second model, human, blocked) and a short list of what must always reach a human: deletions, anything exposed to the outside, the control plane itself, cost and bulk changes.

2. The guard can't live in the agent

Why rules in the agent's own config fail once it can edit that config or read untrusted input, and where to put the gate instead: separate credentials, read and write split into two servers, changes only as a proposal.

3. Approve the plan, not the description

Approvals bound to a plan hash, so what you approved is what runs. Applies to infrastructure plans, pull requests and root scripts.

4. A reviewer model that says no

How a second model checks a request against the reason the agent gave, when it may decide on its own, and how it falls back to you.

5. One message per task, not twenty

Bundling all approvals of a task so you decide once with the full picture, plus the audit log that lets you see later who approved what.

6. Templates

Allow and deny sets for Claude Code, MCP configurations with read/write split, a pull request gate, and checklists for your own review.

Format: a guide (PDF and Markdown, in English and German) plus a template repository. Written from a setup in daily use, not from a vendor deck.

49 €

The playbook is not written yet. I will write it if enough people want it. Join the waitlist and I will email you once when it is ready. No newsletter.

CHECK YOUR OWN SETUP FIRST

The free agent-permission-linter (MIT, Python, no dependencies) reads your Claude Code settings, .mcp.json and the configs of Cursor, VS Code and Claude Desktop, and flags what is too broad:

$ python3 agentlint.py
[HIGH  ] CC003  .claude/settings.json  permissions.allow[0] = 'Bash(python3:*)'
         'python3' with arbitrary arguments runs any code; this is Bash(*) in disguise.

It runs locally and sends nothing anywhere.

WHO WRITES THIS

Tobias Zander. CTO since 2011, writing software since school, and running multi-agent systems inside my own companies and in my homelab. More on tobiaszander.de.