Skip to main content
Taskforge turns real work into RL tasks that are calibrated and verified. A task is a Harbor bundle in your repository: an instruction, an environment, a reference solution and a verifier. Taskforge drafts it, proves the verifier, runs agents on it at k and reads every trajectory. People lead the work. You seed it with a repository, a ticket or a draft. You steer each step and read what it returned. You answer the questions a run asks, and you accept or reject each task. The loop does the drafting, the checking, the running and the reading.

​
The 3C loop

Three nested loops, ordered by cost. A check takes minutes, a run at k takes up to an hour, and a question to a person can take days. Most iteration happens inside; the outer loops run only on what survives. Each step is one tool call. You decide the next one from what the last one returned. run_repo_to_tasks runs the whole loop for you and still brings its questions to you.

​
Where it runs

Taskforge is served from one MCP server, kennet, inside Hum or any MCP client. Kennet is the infra under it: sign-in and tiers, Modal workers, Harbor runs, the model gateway and the run volumes. See Architecture.

​
Next