Tencent Open-Sources LoopForge: An Agent Saying Tests Passed Doesn't Count—Only Exit Codes Do
Tencent has open-sourced LoopForge, a self-hosted multi-agent software delivery platform. It doesn't give agents credit for whether tests were actually run after the code was written; it only looks at script exit codes. But the comments section poured cold water first: if the process self-evolves, who accepts it?
Tencent has open-sourced something called LoopForge, a self-hosted, self-evolving AI software delivery platform. In one sentence: it lets tools like Claude Code and Codex do work in different roles according to the processes your team has standardized.
Let's first lay out the problem clearly. There are now many multi-agent coding solutions, and where they fail often isn't "they can't write it." One commenter put it bluntly: the agent replies "all tests passed," but when you check the logs, the tests never even ran. Code generation capability is increasing, but the reliability of the delivery process hasn't kept up.
LoopForge's solution is to break the process into roles and enforce hard gates at every step.
A bug-fix task is split into five stages:
- Plan Agent analyzes the issue and produces a fix plan
- Code Agent changes the code
- Review Agent reviews independently, so the one who wrote the code doesn't review their own work
- Script Gate actually runs the build, tests, and lint, and determines pass/fail based on exit codes; it does not accept the agent's verbal report
- Human Gate provides human review; work continues only after approval, and if it fails, it is sent back to a specified step for rework
The Script Gate design was repeatedly mentioned in the comments. Some said this is the dividing line between it and a toy: gating on exit codes is more trustworthy than the agent's own account.
The process itself can be reused. An effective development method can be saved as an SOP template, and the next similar task can run with the same role division, inspection rules, and approval flow. The execution status, logs, and artifacts of each node are retained, with support for pausing, resuming, retrying, and rerunning from a specified step.
Two working modes:
Main Bot mode: requirements are submitted in natural language; the system creates a Docker sandbox, calls agents in different roles to do the work, and generates a PR when necessary.
SOP mode: standardized R&D, organizing planning, coding, review, testing, and human approval into a repeatable workflow.
But the comments section wasn't all praise.
One question came up repeatedly: in the self-evolution step, who accepts the changes when the process changes? If it still relies on Human Gate, the speed of evolution is bottlenecked by humans. Another commenter added: if SOPs can be reused, they should also be versioned; process self-improvement should ideally first run a comparison against the old version on shadow tasks before deciding whether to replace it. Otherwise, just as code regressions are brought under control, the acceptance process itself starts to drift.
There are also implementation details. LoopForge is MIT-licensed and self-hostable. But a commenter reminded people to confirm two things before deployment: by default it mounts the host Docker socket, so it cannot be used for multi-tenant isolation and is only suitable for trusted teams; and keys should preferably be configured separately for each person. If these two items aren't verified, don't install it into the delivery process by default.
Tencent's judgment for this project is stated in the material: Coding agents are increasingly capable of writing code, and what truly affects enterprise adoption is the reliability of the software delivery process.
That holds up. But the last link in reliability is still people. Script Gate can silence the claims of "I tested it," but for Human Gate and process evolution itself, there is currently no exit code to look at.
发布时间: 2026-11-17 12:30