AI · AI agents · process runner · human sign-off · Python · Flutter
Crewline
A process runner for AI agents: it runs the critic, reviewers and tests with time and cost limits, stops for a human decision, and the diagram shows the execution trace.
two71.io/projekty/crewline/en/ → (project site)The challenge
An agent ends a task with “done”. To find out how many times the plan came back for a fix, whether the tests passed and what the agent is waiting for, someone has to read a long transcript. Reviews run by the agent's own session used a lot of tokens, and whether a step was green was guessed from text. A company that wants agents to take on part of the work needs a process that runs the reviews itself, keeps the cost in check and waits for a person where it matters.
The solution
The team describes the process from blocks in an editor. The agent's session writes the code and content, and Crewline runs the review and check stages with one command: a critic, several AI reviewers (including a model from another vendor) and tests, in parallel, each with a cost and time limit. Tests, API requests and screenshots run once before the review as shared evidence. The verdict comes back in a fixed JSON format; a bad format, no answer or a hit limit means a send-back with a reason. The “You” step waits for “Approve” or “Send back with a note” from the page or a phone, and the agent's session gets the decision at its next step. On a sample task, input-token usage fell from about 38.7M to 15.7M (−59%); part of that is narrower checking, and with shared evidence we estimate about −50%.
Scope & technology
The whole product, used day to day in two71's own work. A Python server with no external dependencies, running on the team's computer. A stage runner: claude -p with a JSON schema and a cost limit, Codex as a second model, test commands as evidence. Token usage per event, stage and task. A process editor with version history and AI permissions (read-only, propose, edit). A kanban board engine linked to runs. An internet access server that only the team's computer sends data to; processes marked as local never leave it. A Flutter Android app with notifications and decisions from the phone. In progress: running a whole task from request to release.