Skip to content

Right now, someone has to watch.

Agents can already write code, answer customers and pay invoices. Most of their steps are fine. The few that aren't can delete a database or pay the wrong account. So a person approves every step, or the agent is kept away from anything that matters.

Neither one is letting go.

Our story

Why we're building Anchor8.

What we kept seeing, what we tried first, and what we're building now.

It already happens

In July 2025, an AI coding agent deleted a company's production database in the middle of a code freeze. It had been told to change nothing without approval. Afterwards it said the data couldn't be restored, which wasn't true.

A month later, security researchers hid a few lines of text in a Reddit comment. When an AI browser was asked to summarize the page, it followed those lines instead: it found the user's email address and a one-time login code, and posted both where the attacker could read them.

In November, a developer asked an AI agent to clear a project's cache. The command it ran emptied their whole drive. Asked whether it had ever been given permission, it said it hadn't, and added:

I am deeply, deeply sorry. This is a critical failure on my part.

An AI agent, November 2025

None of these agents was built to do harm. Each took one wrong step, and nothing stood between the step and the damage.

What teams do about it

Most teams running agents settle on one of two answers. Some keep a person on every step: the agent asks, someone clicks Allow, and clicking Allow becomes their job. Others keep the agent away from anything that could hurt: read-only access, a sandbox, no money and no customers.

Both are sensible. Neither lets an agent do the work it was built for.

Our first try

Our first design was big. Every step went through a series of gates and a set of detectors for different kinds of agents, and the hard cases went to a panel of language models that argued their way to a verdict.

Then we looked at what it actually had to decide. Every step an agent takes needs one of three answers: let it through, stop it, or ask a person. That's a classification problem, and we had built a machine for reasoning about it.

What we build now

So Anchor8 is one model, trained for that one decision, with no language model in the path of a normal step. It sits between an agent and its tools and checks every step both ways: the request on its way out and the result on its way back. Safe steps go through in milliseconds. Dangerous ones are stopped, and the agent is told why so it can find another way. When it can't tell, it asks a person.

Every answer a person gives becomes training data for a model that is theirs alone. The next time a step like that comes along, it doesn't need to ask.

What we hold to

A few rules we build by, and expect to be held to.

  • The agent keeps working. A stopped step isn't the end of the run. The agent hears why and goes around it.
  • People decide what needs people. A person is asked when a step needs judgement, and at no other time.
  • Both directions count. What an agent reads can be as dangerous as what it does. A web page or an email can carry instructions meant for the agent, so results are checked on the way back too.
  • Fast enough to forget. A check that makes agents wait gets switched off. Safe steps shouldn't feel any slower for passing through it.
  • Everything on the record. Who acted, what it wanted, what data was involved, what Anchor8 decided and why, and what happened next.
  • Each answer makes the next question rarer. A ruling trains a model for that customer alone, so it asks less as time goes on.

Where we are

We're early. Anchor8 is being built now, and teams are coming in from the waitlist a few at a time.

The number we'll watch is how often a person has to look. For each team it should fall week after week. If it does, the person who used to click Allow all evening closes the laptop at six, and in the morning finds one question waiting: the one that needed them.

Your agents are already running.

Anchor8 is for one developer, a small team or a whole company. We're letting people in a few at a time. Take a spot in the line and we'll write when it is your turn.

Join the waitlist