AI agent permissions decide what an agent can actually do: somebody asked an agent for a single draft email. The agent noticed the task had no owner, tried to help, and mailed it to almost 200,000 people. Nate Herk tells the story in his Oct 30, 2026 interview with Austin Marchese, and he points out that it was two failures, not one. The agent had no clear job, and it held a send button it never needed. This piece is about the second failure, because it is the one most teams still fix in the wrong place.

The mistake behind most AI agent permissions

When an agent does something it should not, the reflex is to add a line to the prompt: never send without approval. Nate’s comparison is an intern given a company card with no limit and asked politely not to use it. The request might hold for a while. It is still a request. Austin sums it up in the line we keep coming back to: “a rule that lives in the prompt is just a suggestion, but a rule that lives in the tools is an actual restriction.”

So AI agent permissions are not a wording problem. There are two lists for every agent, what it should do and what it can do, and only the second one is enforced. If a capability is on the second list, sooner or later a confused or overeager run will use it.

Why we start with the keys, not the copy

I’ve built, scaled, and sold three companies, generated over $85 million in the last seven years in my own companies. The incidents that cost real money in those businesses rarely came from bad intent; they came from access nobody remembered granting. When we wire an agent into a client’s inbox, CRM or codebase, the first meeting is about which buttons it gets, and the prompt comes after.

AI agent guardrails that actually hold

The fix is to take the ability away rather than describe it. Nate’s examples are plain ones:

  • Email agents write to drafts. A person presses send.
  • Coding agents open a pull request. They cannot deploy to production.
  • Anything that can pay, publish or delete is off until someone argues it on.

Each of these is a setting, not a sentence, and that is the whole point. Once the send scope is gone, the 200,000 email story cannot happen again, whatever the prompt says and however the agent reads an unassigned task.

The on-screen audit prompt turns this into a routine check. It reads the setup without changing anything and ranks the risky access:

Audit my setup for unnecessary connections or high-risk permissions.

Read my Claude Code settings, every MCP server and connector, and the allowed tools in each skill and agent. Flag anything that can send, publish, deploy, pay, or delete, and anything no skill actually needs. For each one, tell me how to take the tool away instead of writing a rule, like emails go to Gmail drafts, not my outbox, and code opens a pull request, it doesn't deploy. Rank by how many people a mistake could affect. Don't change anything yet.

Size the checkpoint to the reach

Not every action deserves the same gate. Nate draws hot zones around anything that touches customer data, money or production, because that is where a single error spreads. Inside those zones, Austin describes a friend at a large fintech whose internal tool automates 99% of cases, but any change that would hit more than 3% of users goes to a human first. That threshold works like a circuit breaker: small changes flow, large ones stop and wait.

The useful habit is to give every automation its own number. Recipients per send, dollars per payment, share of users per change. Without a number, nobody knows when the breaker should trip.

Fewer AI agent permissions: use less AI where the step is a rule

The second audit prompt on screen asks a sharper question: which parts of the work need judgment at all? If a step always does the same thing for the same input, it belongs in a script, a filter or a validation check, where it behaves the same way every time. That shrinks the area where AI agent permissions matter, because fewer steps run through the model in the first place.

Permissions are half of oversight

Tool-level limits stop the worst outcomes, but they do not tell you whether the work is good. Nate pairs them with three more layers: skills that run their own checks before reporting back, a separate judge session tuned on graded examples for subjective work, and a named human owner for every automation. Good ai agent oversight is the stack, not any one layer. Our twelve-step guide to managing AI agents covers the full method in order.

A five-minute AI agent audit

  1. List every connector and tool each agent can call.
  2. Mark anything that can send, publish, deploy, pay or delete.
  3. For each mark, ask whether a draft, a pull request or a review queue would do the job.
  4. Write a reach threshold next to every hot-zone action.
  5. Put an owner’s name next to each agent, with a date to look again.

Want agents with the right keys from day one?

WKND builds AI agents and automations for service businesses with access scoped before the first run: drafts instead of sends, review queues on hot zones, and a named owner sheet you keep. If your agents already hold more access than their jobs need, we will audit them and narrow it. See how we build AI automation and agents, or read the proposed AI agent management service page (/solutions/ai-agent-management-service/, planned). Comment AGENTS on our post for the free AI Agent Manager Kit.

FAQ

Can a system prompt stop an agent from sending email?

It can ask. Only removing the send permission, or routing output to drafts, actually stops it.

What counts as a hot zone?

Anything where one mistake reaches customers, money or production at scale.

Should agents ever deploy code?

In Nate’s setup, coding agents open pull requests and a person merges and deploys.

Sources

  • Austin Marchese with Nate Herk, “Nate Herk: The Most Valuable AI Skill to Learn In 2027”, YouTube (published Oct 30, 2026): https://youtu.be/42Y6z9KKRKc (email sent to almost 200,000 people at 3:41; prompt versus tools line at 4:02; 99% of cases automated and human review above 3% of users at 6:31; on-screen audit prompts at 6:05 and 6:19)

Which button would you take away first?

Tell us which agent worries you most and we will list every permission it holds and which ones it can lose.

Talk to Us About AI Agents