Grex 2.0 does one thing first: finish real work

Grex lets an operator hand bounded work to AI agents running on computers the operator owns, then come back to a result someone has checked. Grex 2.0 is a new version, rebuilt from scratch this month. It is deliberately narrow. Before Grex does anything broader, it has to show that one loop works reliably.

Why Grex was rebuilt

The first version of Grex grew many parts at once: capability packs, schedules, research pipelines and several kinds of agents. Too often, the hard part was the plumbing between those parts, not the work itself. The evidence for any one outcome was spread across too many places.

Grex 2.0 starts again with less. It is a new codebase with one job to do well.

The loop Grex 2.0 is built around

Every piece of work is a mission. A mission is one written task with a fixed starting point, a list of acceptance criteria, the files it may change and a deadline.

1. A builder agent makes the change on one of the operator’s machines, called a worker.
2. Grex runs fixed checks and keeps everything the builder produced: the patch, the logs and a record of what ran.
3. A different AI reviews the exact change against each criterion. The builder never grades its own work.
4. The operator accepts the result, rejects it, or asks for a bounded correction.

That is the whole product for now. Everything else waits until this loop is dependable.

Human and AI operators get the same controls

Every mission has an operator: the one who writes it, launches it and decides on the result. That operator can be a person or an AI assistant. Grex treats both the same way.

A person works in the dashboard. An AI assistant works through the API, or through MCP, the standard way assistants call tools. Every action has all three: creating an agent, writing and launching a mission, reviewing the transcript and accepting or rejecting a result. Each produces the same record. Grex holds to this as a rule. If an action can be done through the API, it can be done in the dashboard, and the reverse.

This week the dashboard was rebuilt around that rule. Each kind of thing (missions, agents, workers, teams and model routes) now has a list. Each item in a list opens a detail page with the same actions an assistant has. This week’s MiniMax missions were written, launched and accepted by an AI operator working through the API. A person reading the dashboard sees exactly what that operator did.

What the focus means in practice

Your machines and your tools. Workers run on Linux, and Windows is next. Grex uses the AI command-line tools the operator already has, such as OpenAI’s Codex and Anthropic’s Claude Code, and the subscriptions behind them. Grex no longer ships its own copies.

Strong review, cheaper building. The reviewer is a frontier model. The builder can be a cheaper one, sent through what Grex calls a model route. This week, a builder running MiniMax-M3 produced a change that the reviewer accepted on both criteria in under three minutes. That is one mission, not a trend.

The operator can see what happened. Each mission now has a readable transcript: what the agent said, which commands it ran and which files it changed. Secrets are removed before anything is shown.

Feedback that carries forward, under the operator’s control. Agents keep a short list of do-and-don’t lessons. A lesson becomes active only when the operator saves it. The next step is to learn from corrections and to show whether each lesson actually helped.

What Grex 2.0 is not doing yet

• It does not run open-ended goals. Every mission is bounded and reviewed.
• Nearly all the evidence so far comes from coding missions.
• Windows workers are being installed now and have not yet finished a mission.
• Grex cannot yet confirm which model answered when a builder runs through a model route.
• One mission this week was rejected because its acceptance criterion asked for evidence the reviewer could not see. Making that evidence visible to the reviewer is the next fix.

The aim is narrow on purpose. If Grex can reliably turn a bounded task into a reviewed result, it earns the right to take on more.