Autopilot
Autopilot assembles a board of specialists, turns their discussion into a plan with milestones, then works through it in a loop — proposing the next move, doing it, and coming back to you for the call. It runs on your real project, with the model you choose, and it can't mark a milestone done without the artifacts to show for it.
Not a pasted snippet or an uploaded snapshot. Agents open the files you have open, edit your working copy, run the build where the code actually lives, and commit. What you get back is a change on disk, not something to copy out of a chat.
Any OpenAI-compatible endpoint. A hosted provider like Groq or OpenRouter, a model on your own network, or something running on the machine under your desk. Your key, your endpoint, your choice — and changeable per project at any time.
A milestone turns green only when there is something behind it — files written, commands run, builds attempted. "It's done" with nothing to show for it is the characteristic failure of agent systems, so a run is checked for evidence before it gets to report success.
Autopilot doesn't answer once and stop. It reviews the milestone spine, works the next one, checks what actually landed, and comes back with the result before carrying on — so a long piece of work arrives as a sequence of small reviewable steps rather than one opaque run you either accept whole or throw away.
Nothing appears behind your back, either. Agents and groups are proposed for your approval before they are created, and you get artifacts at the end rather than a description of them.
Agent groups put several specialists on one goal and let them work it out in the open. You can read the exchange as it happens, see artifacts appear on the canvas beside the conversation, and step in whenever you want.
Long discussions don't stay transcripts — they're synthesised into decisions, milestones and plans you can act on.
Point WorkByte at a repository and it builds a semantic index and a graph of the code. Ask "where do we handle expired sessions?" and get an answer even when nothing is named that.
Agents get the same operations your IDE has — go to definition, find references, rename — plus git end to end: clone, branch, commit, push, open a pull request.
Agents read, understand and edit a real repository, with language-server-grade navigation rather than string search — then commit and open a pull request.
Turn a goal into a plan, a plan into milestones, and let a group of specialists work them — checking back with you between steps.
Point it at something unfamiliar and ask questions in English. Answers come from a semantic index and a code graph, not a keyword match.
Grounded answers with citations back to the source, plus a learn mode that explains a topic instead of just answering.
Documents, diagrams, media and exports — produced by skills that run real scripts in a sandbox, on your machine or in the cloud.
Name it, instruct it, choose its model, give it tools. Agents here are data rather than code, so the ones you make are the same kind of thing as the ones that ship.
The free tier runs the sandbox on your own machine — so your work stays with you, and there's nothing to meter.
