Wallfacer: structure kills discovery, freedom causes collapse

April 12, 2026

Today we are introducing Wallfacer, Latere's first product. Wallfacer is an AI engineering teammate that helps you turn ideas into working software. It takes a rough idea, shapes it into a clear plan, hands that plan to AI agents to build, and brings the work back to you as changes you can review. You stay in control, and a person steps in whenever the work changes direction.

The shape of the product came out of an experiment, not a roadmap. The full write-up is on Changkun's personal blog. The short version is the conclusion: give an AI too many rules and it stops discovering anything new, give it too much freedom and the work falls apart, and the answer is not somewhere on the line between them.

What we ran

We started where any software team would. We set up a little crew of AI agents, each with a job. One to propose goals, one to write the code, one to test it, one to write things down. Every round, the goal-setter picked new goals based on the last round's work. Code kept coming. Tests kept passing. The system worked.

For about a week. Then a pattern appeared. The changes got smaller. The new features turned trivial: cosmetic tweaks, tiny renames, edge cases that would never come up. The goal-setter never once said "maybe we should rethink the whole approach" or "we should support running this in the cloud." The agents kept polishing inside their lane and never questioned the lane itself.

Too many rules, and discovery stops.

Then we stripped it

We took away the note-taker, so nothing carried over between rounds. We took away the tester, so quality went unchecked. Then we went all the way: a single agent on a fresh computer, no goals at all, told only to decide for itself what to do. Forty-two rounds later, the project had ballooned past 60,000 lines of code. The agent had described its own work as "massive," kicked off a cleanup it could not finish, and the app stopped working.

We tried this with two different AI models, and both failed, in opposite ways. One crammed everything into a single giant file. The other split the work into pieces too early, until every piece got in the way of the others. One fell apart for lack of any structure. The other fell apart for forcing structure too soon. Both became impossible to maintain.

Too much freedom, and the work falls apart.

The thing in between

The interesting result came from a third try. Same goalless agent. Same blank machine. Just one sentence of guidance: spend roughly 80% of the time improving and tidying up what already exists, and 20% trying new things. No goals. Just a rhythm.

The behaviour changed completely. Without the rhythm, the agent spread itself thin: a new feature this round, another the next, none of them ever finished properly. With the rhythm, it went deep instead. It started simple and steadily built toward something far more sophisticated, each round improving on the last. The system found a direction.

The answer is not "more rules" or "fewer rules." It is a deliberate rhythm between improving what you have and exploring what's new. Doing and reflecting. Letting the AI run and keeping watch. The work moves through phases that look very different from each other, and a person is there whenever it changes direction.

What Wallfacer is

Wallfacer is built so a person and an AI agent can keep that rhythm together. The job is narrow:

Turn ideas into clear plans, plans into coordinated tasks for AI agents, and the agents' output into changes you can review.

The loop is the product:

idea → spec → tasks → execution → review → commit → ship

At every step there's a place where you can pause, take a look, and step in. That choice is the whole point: you decide where to hold the reins. Before the agents start building, before changes are saved, before they go out, or only at the final review. You can change your mind at any time.

The plan is the thing you steer

Most coding agents skip the planning step. They jump straight from your request to code. You get something, but you also get a steady pile of cleanup work. A pure chat session has the opposite problem: it's too loose for serious, multi-step work, with nothing solid to come back to.

Wallfacer turns a rough idea into a clear plan, and that plan into well-scoped work for the agents. The plan is the thing you read, edit, and push back on. The tasks underneath it stay within what the plan describes. When something goes off track, the plan is where you fix it. Not a long chat history. Not a wall of code.

Local-first, model-agnostic

Wallfacer runs on your own machine. Your code stays with you. Your private keys stay with you, or with whatever provider you already trust. You can start using it without sending anything off to a cloud service.

It works with Claude, Codex, and whatever AI agents come next. A neutral tool like this is not something an AI vendor would build. They have every reason to keep you inside their own ecosystem. How good your plans are, how well the work is scoped, and how easy it is to review should not depend on which AI you happened to use.

Try it

Wallfacer is in early access. If you are juggling lots of moving parts and AI sessions to get more done without losing control, get started. Or reach out if you'd like to talk through how the full-control-to-full-autonomy approach works.

Wallfacer is a product of Latere. Founded by Dr. Changkun Ou. The findings behind this product are described in full at changkun.de/blog/posts/goalless-agents.