Replying to @_p18z @TencentAI_News
Thanks for the question. LoopForge Gates support more than simple test pass/fail checks. Teams can configure a diff coverage tool as a Command Gate to measure coverage of changed lines. If coverage falls below the configured threshold, the command returns a non-zero exit code, blocking the workflow or triggering rework.
7
Replying to @brianmcgrath @TencentAI_News
Thank you for the question. LoopForge evaluates runs using the scoring policy configured for the project. Changing a Gate alone does not alter that policy. Teams can use the version history for a role or SOP to compare saved revisions with the current definition and restore an earlier version when needed.
1
14
Replying to @aakashpawar1999 @TencentAI_News
Thank you for the question. For example, if a role’s Loop limit is set to three iterations, the Loop can run at most three times, including the initial attempt. When the workflow reaches a Gate configured with a test command, LoopForge runs the actual test. If the test fails and the Gate is configured for self-correction, the error details and test output are passed back to the Agent as context for the next iteration.
If the same test still fails on the third iteration, automatic correction stops. Depending on the configured policy, the node is either marked as failed or paused for human review. Teams can inspect the tests executed in each iteration, their output, and the failure reasons in the run history.
1
25
Replying to @dusangran @TencentAI_News
Thank you for the feedback. LoopForge’s pipeline does not automatically rewrite SOPs. Every run preserves its execution trace and artifacts, and the AI only proposes improvements based on the results. Any SOP change must be reviewed and approved by the user before it takes effect, with version comparison and rollback supported.
With the planned RSI system, candidate versions will first be tested against the current version in an isolated environment, then evaluated on held-out cases. A candidate will only be published when the evidence meets predefined thresholds and the user approves it. The relevant versions and validation records will be retained for comparison and rollback.
16
Replying to @agneau123 @TencentAI_News
Thank you for the feedback. By design, improvement recommendations generated by LoopForge’s AI evaluation never automatically overwrite an existing Review Gate. Changes to roles or workflows are versioned and shown as a clear diff. They only take effect after approval by the designated owner, and previous versions can be restored when needed.
31
Replying to @h0ick @TencentAI_News
Thank you for the feedback. Building on SOPs and other foundational assets, we plan to introduce RSI, or recursive self-improvement. Candidate versions will first run in an isolated environment, be compared with the current version, and then be evaluated on held-out cases that were not seen during the improvement process.
During review, teams will see what changed, why it changed, which tasks improved or regressed, and the supporting run evidence. If the evidence is insufficient, the current version remains in place. A candidate will not be published simply because the AI says it passed.
By default, the designated owner still decides whether to adopt a change. Automated trials, publishing, and rollback require explicit authorization and must satisfy predefined thresholds. After release, LoopForge will continue monitoring the new version while retaining the previous version for rollback.
Our goal is to reduce the manual cost of repeated trial and error while making every change understandable, verifiable, and reversible.
1
42
Replying to @0xnoobC @TencentAI_News
Thanks for the question. Normally, no human intervention is needed. When the review agent explicitly submits a rework decision, LoopForge records the reason as feedback on the code node and automatically routes the workflow back. The coding agent makes the changes in the same workspace, then the work goes through review and testing again.
Rework has a configured limit, so the two agents cannot loop indefinitely. Once that limit is reached, the workflow pauses for human review. The owner can accept the current result, add context and retry, or adjust the SOP. Each review decision, rework reason, and node state is recorded, so the owner only needs to step in when the escalation condition is reached.
1
1
35
Replying to @TencentAI_News
想自己跑一遍,入口放这里👇
官网:loopforge.cn/
GitHub:github.com/Tencent/LoopForge
欢迎体验,也欢迎来提 Issue。
64
LoopForge 把团队的工作流程沉淀成 Loop,在云端 7×24 小时持续运行。
任务状态和执行进度全程可视化,全队随时都能看见。
产品、研发、测试共用一个会话,讨论、代码和产物都沉淀在同一份上下文里。
一个会话,接住整个团队。
1
5
146
Replying to @ajs6888
感谢您的建议。目前 LoopForge 面向可信团队共用实例,登录用户可查看实例内的项目、需求和运行记录,项目权限主要限制修改和执行。每次 SOP 运行使用独立 Docker 沙箱。后续会考虑项目间用执行机隔离、项目内用 Docker 隔离,进一步完善项目边界。欢迎继续体验和反馈。
12
Replying to @huishu96225748 @NFT_Chen
谢谢您的建议。我们会保持努力优化,继续提升。目前LoopForge 里,写代码和审代码的不是同一个角色。代码写完后,会交给专门的评审角色检查;发现问题就带着具体意见退回修改,改完再审。不会让写代码的自己给自己打分。
1
18
团队里每个人都在用 AI,交付却没变快?
那是使用姿势不对。代码写得是快了,可从“写完”到“能交”,还是得一遍遍对需求、盯评审、催修改。
我们开源了 LoopForge:把每个人的经验配成 Agent 角色,把团队流程写成可迭代的 SOP,在云端 7×24 跑。不合格自动打回,人只管验收。
MIT 开源,一键部署👇 github.com/Tencent/LoopForge
2
5
191
Replying to @CY160909
Really appreciate the shoutout! 🙌 We’re building LoopForge to turn team delivery playbooks into agent-executable SOPs—from requirements and design to code review, testing, and release. It’s open source, self-hosted, and evolving fast. Feedback is always welcome!
1
1
63
🚀 Let your agents team up and work in the cloud. We just open-sourced LoopForge, a self-evolving cloud platform for software R&D. LoopForge moves the work to the cloud:
☁️ 24/7 cloud workspaces: close your laptop, your agents keep running
🤝 One shared session for the whole team: PMs, designers, QA and engineers take turns talking to the AI, and the context carries over automatically
🔁 Autonomous loops in the cloud: scheduled patrols find issues and fix tickets on their own, so nobody has to babysit on-call
🧬 Self-improving workflows (RSI): what works gets captured as an SOP automatically, so it gets better the more you use it It already powers 20+ internal teams at Tencent and has shipped 200+ requirements end-to-end, unattended, with zero incidents. Come take a look and tell us what you think. Code was never the bottleneck. The handoff is. LoopForge picks up the baton between agents that nobody was there to catch.
👉 Website: loopforge.cn
👉 GitHub: github.com/Tencent/LoopForge (a ⭐ means a lot!)

