Wink Pings

Tencent Open-Sources LoopForge: Multiple Agents Queue Up to Work, but the Final Step Still Needs Human Approval

Tencent open-sourced a self-hosted Coding Agent workbench. It strings design, implementation, review, testing, and human acceptance into a repeatable SOP, supports ten CLIs, runs in Docker sandboxes, and can roll back to a specified node to rerun on failure. Someone in the comments did the math: splitting the same review into four steps raised tokens from 960k to 2.23M. The process layer is free, but compute and approval have become the new bottlenecks.

LoopForge

Tencent has open-sourced LoopForge, a software development and delivery workbench for Coding Agents. The problem it aims to solve is like asking an Agent to fix a bug: run it ten times and you get ten different results. Sometimes it skips planning, sometimes it forgets to run tests, sometimes it reviews its own code, and sometimes it stops halfway with no one noticing. That's enough for a demo, but not enough for delivery.

LoopForge's approach is to put every delivery through the same pipeline. It breaks design, implementation, review, testing, and acceptance into nodes and organizes them into an SOP (standard operating procedure), so every requirement follows the same rhythm.

What nodes are in an SOP?

SOP 流程

Nodes fall into four categories.

Agent node is a role that packages prompts, executors, models, rules, Skills, and MCP tools. Design, implementation, review, and testing each get their own independent role.

Script node is a deterministic shell step used to build, run tests, lint, and generate reports. Where AI isn't needed, don't use AI.

Human gate stops the process and waits for a person to inspect the artifact and decide whether to approve or send it back.

Sub-SOP calls one SOP as a subflow. There is also a preview build that builds a static site and archives it for review.

The key is what happens after a failure. If a step fails, the process can roll back to a specified node and rerun, preserving node status, tool-call records, logs, and snapshots of each step's artifacts. Someone in the community noted that rolling back to a specified node is more critical than logs. Logs only record; rollback means the process can recover.

Four SOPs are built in: backend delivery, frontend delivery, product proposal, and bug fixing. Built-in templates are read-only blueprints; after copying them into a project, you can change roles, rules, and environments, and then publish a version.

内置 SOP 模板

Two ways to work

The first is chat. You describe a requirement in the main Bot, and it spins up a Docker sandbox on demand, pulls your repository into it, calls project roles as sub-Agents, and finally delivers a PR. If a reply is halfway through, you can cut in and change direction; if the process is halfway through, you can pause, resume, or take over.

聊天工作台

The second is running an SOP. The engine executes Agent, Script, Human gate, and sub-SOP nodes in dependency order. Each node runs in its own sandbox and delivers artifacts and a PR when it finishes.

SOP 设计器

If you don't want to set up a server yourself, there is also a Local version that can be installed into an existing CLI with one command:

npx -y loopforge-cli@latest install claude

Then run /start-devflow "你的需求" in the target project. Claude Code, CodeBuddy, Codex, and Cursor support the full and portable versions; Pi, DeepSeek Harness, and Hermes Agent use the portable version.

On the platform side, ten executors are supported: CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness, and ZCode. Each SOP role can independently choose its executor and model.

人工审批

Where the math matters

Process orchestration sounds like all upside, but every step is an inference call. Someone in the comments measured the cost: for the same code review, a single Opus skill used 960k tokens; a review workflow split into four steps used 2.23M tokens, with the review step alone accounting for 790k, while total elapsed time was just over 3 minutes in both cases. Before adopting LoopForge, it's worth running it on a real requirement and comparing the bill.

Another reader put it more bluntly: open-source the orchestration layer and leave compute demand to the underlying layers—this is the same playbook as various vendors releasing their own Agent systems around the same time. The process is free; the bill lands on tokens.

There is also a risk that isn't about compute but about people. Someone in the comments asked a very practical question: multi-Agent division of labor is the right direction, but in practice the most common failure is that acceptance criteria aren't defined clearly enough. The design, implementation, and review Agents each say their own thing, and ultimately a human still has to backstop it. The human gate is a design highlight of LoopForge, but whether there are enough people to approve and whether the criteria are written clearly determine whether it is a control point or a new bottleneck.

Two things to check before deployment

In the default Compose installation, both SOP runs and session workspaces start sandbox containers through the host Docker socket. Docker socket permissions are equivalent to controlling the host daemon; this configuration is for trusted teams, not a multi-tenant isolation solution.

New installations listen on 0.0.0.0:8090 by default. Before the admin account is created, restrict access to your own IP, and configure an HTTPS reverse proxy before exposing it externally.

Quick start:

git clone https://github.com/Tencent/LoopForge.git
cd LoopForge
./start.sh

After installation, open the address printed by the installer; the first registered account automatically becomes the administrator. Also, you can run a demo project without an API key: an SOP with only a script and a human gate. The script writes a report and then stops at the approval step, waiting for you to click approve.

Project address: https://github.com/Tencent/LoopForge

MIT license, self-hosted. It does not replace CLIs such as Claude Code and Codex; it manages the layer beyond those CLIs: who does what, in which sandbox, in what order, which checks to pass, and who signs off.

发布时间: 2026-11-18 01:35