Tencent has open-sourced LoopForge, a self-hosted workspace that runs coding agents as a delivery pipeline rather than a chat toy. The pitch is blunt: ask an agent to “fix this bug” and the result depends on that run. LoopForge’s answer is an SOP, a standard workflow of agent roles, deterministic scripts and human approval gates, executed on your own infrastructure under the MIT license. The repo ships ten coding-agent CLIs as interchangeable executors, including CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness and ZCode.

The interesting part is not that another agent orchestrator exists. It is who built it and what the design assumes. Tencent is not shipping a frontier coding model to compete with OpenAI or Anthropic here. It is shipping the layer above them, and treating every one of those vendors’ CLIs as a swappable part. That is a bet that the workflow, the sandbox and the audit trail matter more than which model writes the diff.

The SOP is the product

LoopForge’s core abstraction is a workflow graph. An SOP is built from five node types: Agent, Script, Human gate, Sub-SOP and Preview build. Each node declares its artifacts and dependencies, and the engine starts a step once its dependencies complete. Agent nodes are roles with their own prompt, executor, model, rules, Skills and MCP tools. Planning, implementation, review and testing each get a separate role, so the agent that writes code is not the agent that reviews it. Built-in SOPs cover backend, frontend, bug-fix and product work.

The Script node is the part worth pausing on. It runs a deterministic shell step, build, test, lint, generate a report, with no model in the loop. The design concedes something the agent-hype cycle tends to bury: not every step needs inference. A test suite either passes or it does not. LoopForge puts the model where judgment helps and a shell command where it does not.

The Human gate node pauses a run until someone reads the artifacts and approves, or sends the work back to a chosen step. Tencent documented a real run: four CodeBuddy roles planned a Python fix, implemented it, reviewed it independently and summarized the delivery. Three tests passed, the run waited for human approval, then verified the accepted code had not changed. That last check is the tell. The system re-verifies the artifact after sign-off, which is the kind of paranoia you build when you have watched an agent quietly drift.

Your stack, their sandbox

The deployment model is deliberately boring. ./start.sh pulls a pinned GHCR image, creates MySQL and the platform services, generates secrets and installs starter SOPs. The service listens on 0.0.0.0:8090 by default. No host Go or Node install is required. You bring model API keys and, optionally, GitHub OAuth.

Tencent is upfront about the security boundary, and it is narrower than the marketing might suggest. In the default Compose install, SOP runs and conversation workspaces both use Docker sandbox containers started through the host Docker socket. The repo says plainly that socket access grants control of the host daemon, so the setup is for trusted teams, not untrusted multi-tenant isolation. New installs listen on all interfaces, signed-in users can read project data across the instance, and teammates can sign up from the login page unless AGENTFLOW_ALLOW_SIGNUP=false is set. This is a tool for a team behind a VPN, not a public service.

There is also a mirror escape hatch. If ghcr.io stalls, you set MIRROR to a registry host that can stand in, and the repo names ghcr.nju.edu.cn as an example. The docs warn that setting MIRROR trusts that host, because the download gets tagged with the official name, and they tell you to check the digest against the GitHub Release. That is a candid acknowledgment of the supply-chain surface a self-hosted tool inherits.

The model-agnostic bet

The most consequential design choice is that every SOP role picks its executor independently, and the built-in roles default to CodeBuddy, Tencent’s own coding assistant. That default is the tell about intent. Tencent built LoopForge to make CodeBuddy first-class, then opened the door to Codex, Claude Code and the rest. The seven agent CLIs after Claude Code keep their model and MCP configuration in a private home directory per run, so each uses the responsible user’s own key and Base URL. When you roll this out to a team, every member configures their own credentials.

Tencent is not competing with the frontier labs on model quality. It is competing on the scaffolding around them, and it is giving that scaffolding away.

That framing has real implications for the AI economy. If the orchestration layer is open source and model-agnostic, the moat shifts toward whoever owns the workflow library, the evaluation record and the integration surface, not whoever owns the weights. A team can swap Claude Code for Codex in a single role without rewriting the SOP. The switching cost lives in the SOPs and the audit history, which is exactly where Tencent wants it.

There are gaps. The repo notes that requirement clarification, AI notification summaries, Bot replies and run evals currently use CodeBuddy, so their model pickers list CodeBuddy models only. Cursor CLI and Antigravity are installed by a provider installer on first chat use and are not installed into SOP sandboxes. ZCode is not bundled at all and must be on the PATH. The executor matrix is broad but uneven, and the CodeBuddy gravity in the non-SOP paths is a soft lock-in worth watching.

What it means for builders

For teams already running agents in production, LoopForge’s contribution is the vocabulary. Separating planning, implementation, review and testing into distinct roles with distinct models is a pattern more shops should copy, and putting a deterministic script between agent steps is the cheapest reliability win available. The human gate is not a concession to distrust. It is a checkpoint that lets a run resume, retry a failed step or rerun from any node without starting over.

The open question is whether an MIT-licensed orchestrator from a Chinese platform giant gets adopted by Western engineering teams that are already wary of self-hosting and already standardized on one vendor’s agent. The ten-executor roster suggests Tencent expects the answer to be yes, or at least expects CodeBuddy to ride along wherever it is not. The repo’s own validation record is honest about scope and known gaps, which is more than most agent launches manage. Watch whether the SOP library grows beyond the four built-ins, and whether the CodeBuddy default in the non-SOP paths gets loosened. Those two signals will say whether LoopForge is a real neutral layer or a Trojan horse with a friendly license.