We wanted AI to do more work.
Instead, we got more work to manage.

We were already using agents every day. They could research, write, analyze, and build faster than we could alone. So we opened more threads and gave them more work.

It looked like leverage. It did not feel like it.

Each new thread became another place to explain the job, check the progress, correct the course, find the output, and reconstruct the context later. The agents multiplied. So did the work of managing them.

The agent was faster.
We were busier.

We stopped trying to manage more chats.

We asked a different question: what would it take to give an agent real work, step away, and return to something we could inspect?

A better prompt was not enough. Repeatable work needed a shape. It needed clear stages, the right context, a definition of done, routes for work that fell short, and a clear place for human judgment.

Once the method was explicit, the interface started to feel different. The unit of work was no longer a chat we had to babysit. It was a run with a record.

We could see what was moving, what had stopped, what needed another pass, and what needed us. We cared less about every tool call. We cared more about the evidence and the result.

That experiment became Slipform.

Slipform is a Mac app for building repeatable agent workflows around how you work. Start with your own method or a useful structure. Codex executes the work today. Slipform keeps the stages, artifacts, evaluations, routes, and decisions with each run.

When a run reaches a human gate, you approve it, send it back, or stop it. Slipform records the decision. You decide what changes before the next workflow version.

We are still early.

The current beta works with Codex on Mac. You create the workflows, approve the work, and own every revision. Broader runtime support is part of the direction, not something we ship today.

We are building toward a workday where more agent work can move at once without creating more work to supervise. A workday organized around evidence, exceptions, and decisions instead of a wall of active chats.

That future is not finished. We are trying to reach it by doing real work, studying real runs, and improving the system with the people who feel this problem too.

Inspect one current-beta workflow and run record →