AI slop is what we fix.

An ecosystem for AI agents: every step visible, every step checked.

↓ 01 · the problem

// 01 · the problem

Code is art a commodity

Teams can generate changes faster than they can check them. When more code is merged than anyone can read, no one really knows what shipped.

Anthropic 2,500 → 30,000+

Merged code changes a day: "around 2,500" in May 2026, "over 30,000" by September.

Redesigning source control at Anthropic, 2026-10-13

Ten major open-source repos 14% → 30%

Pull requests opened July to September and not merged by October 30, median repo, 2024 vs 2026: still waiting or closed. Up in all ten repos, as PRs opened per day went from 16 to 30. PostHog, Grafana, n8n, VS Code, Next.js, Kubernetes, LangChain, Transformers, Supabase, Langfuse. Automation bots excluded; coding agents kept.

Our own GitHub search, 2026-10-30

Research 84%

of 33,596 agent-written pull requests on GitHub got no recorded human review. Human-written ones fare little better. Researchers now argue code review is becoming a rubber stamp.

Research: Duma et al. 2026, section 5.1 · BaxBench, Vero et al. 2025 · SWE-bench patches, Wang et al. 2025 · The End of Code Review, Monperrus 2026

Agents now produce more than people can check.

// 02 · the solution

Every step you can see.

A demo of how it works, on a resume task and a video task.

01

You see every step

A task splits into steps. Each step shows what it made, which model made it, and which model checked it.

02

A different model checks it

Any model can do a step, and any model can check it. In a six-day sample of our own work, another vendor's model found a serious issue in 20 of 25 code changes.

03

Redo any step

Change a step and redo it. The steps that depend on it update, and you don't start over.

Every check leaves a receipt.

// 03 · our advantage

Experience compounds.

We start with teams producing more agent-written code than their experienced reviewers can check. Their accumulated experience is the asset.

Start narrow · Use it ourselves · Let any app plug in

01

Your record stays

It grows with every piece of work, whichever model did it.

02

Across model vendors

Works with models from any vendor. No single vendor holds your record.

03

Thousands of errors caught

We use it every day on our own work.

The Loretoad checks a lesson in its notebook, then catches a bug with its tongue: caught one!

Reviews with past findings caught 5 of 6 reintroduced bugs; without them, 2 of 6.

// 04 · roadmap

What comes next.

  1. Working demotoday
    Agents from different vendors, a shared work record, and durable pipelines.
  2. Domain validatorsplanned
    Open-weight models in harnesses built for specific kinds of work.
  3. AI-native appsplanned
    Apps built for AI: a clear machine contract, every change can be undone, and a machine can read the results.
  4. Open ecosystemdirection
    Any app from any developer and any model from any vendor can plug in.

Domain validators and AI-native apps are in development.