Tencent LoopForge Turns a Good Agent Run Into a Repeatable SOP
Ask a coding agent to fix a bug and the result depends on that particular run. Maybe it planned, maybe it skipped the tests, maybe it reviewed its own code and called it done. Tencent's open-source answer is LoopForge, a self-hosted workspace that jumped from 101 to 484 stars a day on GitHub Trending this Sunday (about 1,420 total). Repo: https://github.com/Tencent/LoopForge
It works in two modes. In chat, you describe the job to a Main Bot, which opens a Docker sandbox, checks out your repos, hands pieces to project roles as subagents, and opens the PR. When a way of working proves itself, you save it as an SOP: a fixed sequence of agent nodes, script nodes, human approval gates and sub-SOPs. Planning, implementation, review and testing each get their own prompt, model, rules, Skills and MCP tools. At a gate you read the report, approve, or send the work back to a chosen step, and you can rerun from any node. It does not ship its own model. Ten coding-agent CLIs do the work, including CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI and Kimi Code.
The latest release, v0.3.9-rc.1, landed Nov 13, before this window, so the news here is the trending jump, not a launch. The release notes are practical: install a missing CLI straight from a failed conversation, per-host SSH keys, agent CLI config writes confined to the home directory, and a pinned image digest with an explicit warning that a mirror host is trusted. The README says MIT; GitHub's license detector reports none, so check the LICENSE file yourself before adopting it.
The interesting part is the unit of reuse. Claude Code's Code Mode and Antigravity's Dynamic Workflows made the script that spawns agents the thing you save. LoopForge makes it a delivery process with named humans in it. That is closer to how a team actually ships, and it says something about what big companies want from agents: not a smarter single run, but the same run every time with a sign-off. The same instinct shows up in plannotator (https://github.com/backnotprop/plannotator), a local review layer for agent plans and diffs, trending for a third straight day at 860 stars a day. The human gate is becoming a product category.
← Back to all articles
It works in two modes. In chat, you describe the job to a Main Bot, which opens a Docker sandbox, checks out your repos, hands pieces to project roles as subagents, and opens the PR. When a way of working proves itself, you save it as an SOP: a fixed sequence of agent nodes, script nodes, human approval gates and sub-SOPs. Planning, implementation, review and testing each get their own prompt, model, rules, Skills and MCP tools. At a gate you read the report, approve, or send the work back to a chosen step, and you can rerun from any node. It does not ship its own model. Ten coding-agent CLIs do the work, including CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI and Kimi Code.
The latest release, v0.3.9-rc.1, landed Nov 13, before this window, so the news here is the trending jump, not a launch. The release notes are practical: install a missing CLI straight from a failed conversation, per-host SSH keys, agent CLI config writes confined to the home directory, and a pinned image digest with an explicit warning that a mirror host is trusted. The README says MIT; GitHub's license detector reports none, so check the LICENSE file yourself before adopting it.
The interesting part is the unit of reuse. Claude Code's Code Mode and Antigravity's Dynamic Workflows made the script that spawns agents the thing you save. LoopForge makes it a delivery process with named humans in it. That is closer to how a team actually ships, and it says something about what big companies want from agents: not a smarter single run, but the same run every time with a sign-off. The same instinct shows up in plannotator (https://github.com/backnotprop/plannotator), a local review layer for agent plans and diffs, trending for a third straight day at 860 stars a day. The human gate is becoming a product category.
Comments