Long AI-agent work is starting to be described as a loop, but a prompt that says "keep going" is not enough for real software projects.

FlowPilot makes the loop explicit and turns a single AI run into a role-based project team. It can plan a route, dispatch background agents into bounded roles like PM, worker, reviewer, and FlowGuard operator, then bring their results back through sealed packets, role-separated review, repair gates, and completion evidence.

The important part is that the loop is model-first. FlowGuard-style process simulation checks the route before the run treats it as safe. During the project, FlowPilot can replan when requirements change, evidence becomes stale, or a check fails. The agent is not just repeating "continue"; it is working through a route it can inspect, challenge, and update.

The practical startup flow is intentionally simple:

1. Open an AI-agent interface that can use local tools.

2. Turn on goal, target, or objective mode if the host has one.

3. Ask the agent to formally start FlowPilot and keep using FlowPilot until terminal return, user input, or explicit stop permission.

4. Wait for the startup window.

5. Enter the project goal, constraints, and expected outcome.

6. Let FlowPilot control the run.

FlowPilot is open-source and local. The long-term idea is not a black-box hosted agent, but an inspectable project-control runtime that people can change, extend, or replace for their own workflows.

I would love feedback on three things:

1. Does "open-source control loop for AI agents" make the product clear?

2. Is the role-based background-agent workflow understandable enough?

3. Which part of long AI-agent work do you most want a runtime to control?

Repository:

https://github.com/liuyingxuvka/FlowPilot