What we’re working on.

The more an agent can do, the more this matters. An agent that edits files, runs commands, and reaches the internet is useful because its actions have real consequences — and those are exactly the things a person should stay in charge of.

So we work on where the line sits: permissions that treat reading differently from changing, actions that stop and ask because of what they touch, and defaults that ask when in doubt. The goal is autonomy you hand over on purpose, not autonomy you discover after the fact.

Most AI tools live one message at a time. Real work doesn’t — it has a goal, a series of steps, half-finished states, and things that go wrong in the middle. An agent is only useful for real work if it can keep track of all of that at once.

So that’s what we build: plans that survive from one step to the next — and one session to the next — a task list the agent keeps and you can read at a glance, recovery when a step fails, and the ability to stop an agent mid-task without losing what it has already done.

When a person does work for you, you can ask what they did and why. An agent should have to clear the same bar — otherwise you’re not trusting it, you’re taking it on faith. Most of that gap is an interface problem, and that’s how we treat it.

In practice, that means work you can actually review: file changes as diffs you read before they land, commands you see before they run, a plan that accounts for every step taken. Nothing our agents do should be invisible to the person they work for.

New models come out all the time, each with different strengths, costs, and weak spots. We build around that fact instead of fighting it: the models are parts of the system, not the system itself — and software built that way outlasts software tied to a single vendor.

We’re deliberately independent of every model provider. Our work is in choosing well: the model a task actually calls for, how long it should think, and the same rules — the same approvals, the same visibility — no matter which model is doing the work.