The Pneumetron
NEWSTOOLSINFRASTRUCTURE
Pneumetron.Tencent releases LoopForge to orchestrate coding agents in self-hosted pipelines
Share
Skip to article content
  1. Home
  2. ›
  3. News
  4. ›
  5. AI Research
  6. ›
  7. Tencent releases LoopForge to orchestrate coding agents in self-hosted pipelines
AI Research·November 19, 2026

Tencent releases LoopForge to orchestrate coding agents in self-hosted pipelines

BY PNEUMETRON|4 MIN READ · 944 WORDS4 MIN READ
Tools
Share
Illustrative photo: internet, information, service, system
Illustrative image: cookieone / Pixabay

Tencent has open-sourced LoopForge, a self-hosted orchestration layer that wraps ten coding-agent CLIs into repeatable, auditable delivery pipelines. The system enforces role separation, Docker-sandboxed execution, and human approval gates to move agents beyond conversational demos.

Key Takeaways

  • 01Tencent open-sourced LoopForge, a self-hosted orchestration layer that wraps ten coding-agent CLIs into repeatable delivery pipelines.
  • 02The system enforces role separation by routing planning, implementation, review, and testing through independent agent nodes with distinct prompts and models.
  • 03Default sandboxing requires Docker socket access, granting host daemon control and restricting deployment to trusted internal teams.
  • 04Token cost accounting and requirement tracking link chat interactions and automated runs directly to PR status and per-node artifacts.

Tencent has open-sourced LoopForge, a self-hosted orchestration layer designed to move coding agents from conversational experiments into auditable, repeatable software delivery pipelines. Published on GitHub under the MIT license, the system wraps ten distinct coding-agent command-line interfaces into a unified workspace that enforces role separation, Docker-sandboxed execution, and explicit human approval gates. The release targets engineering teams that have outgrown ad-hoc agent interactions and need deterministic delivery stages with full traceability.

The core of LoopForge is its standard operating procedure (SOP) engine, which replaces open-ended chat sessions with a directed graph of execution nodes. Each run follows a predefined workflow where agents, deterministic scripts, and human reviewers take turns. The engine supports four primary node types: an Agent node that delegates work to a specific role, a Script node for deterministic shell commands like linting or testing, a Human gate that pauses execution until an engineer reviews artifacts, and a Sub-SOP node that nests another workflow inside the current run. Failed checks can route work back to earlier steps, and every node declares its artifacts and dependencies so the engine only starts a step once its prerequisites complete.

This structure forces a separation of concerns that standard agent chats rarely achieve. Planning, implementation, review, and testing each get their own prompt, executor, model, rules, and MCP tool configuration. A role functions as an SOP node or as a subagent in a live chat conversation. When a requirement enters the system, the engine opens a Docker sandbox, checks out the relevant repositories, and delegates tasks to project roles. The system tracks requirements end-to-end, linking every change made in chat or by an SOP back to the original request. It also monitors PR and MR status, detects merges, and supports direct branch merges when teams prefer to bypass pull requests.

LoopForge does not build its own coding agents. Instead, it acts as an orchestrator for existing CLI tools. The platform supports ten executors: CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness, and ZCode. Teams can assign different executors to different SOP roles or switch them per conversation. The built-in roles default to CodeBuddy, but the system refuses to route a non-CodeBuddy model to a role that strictly requires it, such as requirement clarification or bot replies. Scheduled assistant tasks run on whichever CLI the task specifies.

Credential handling varies significantly across the supported executors, which matters for teams managing shared infrastructure. CodeBuddy accepts personal keys or administrator-configured shared keys. Codex and Claude Code rely on personal keys or local machine logins. OpenCode, Kimi Code, and Pi require personal keys and install from npm on first use, caching the binaries afterward. DeepSeek Harness and ZCode require personal keys only, with ZCode needing manual installation on the service PATH. Cursor CLI uses a provider installer on first chat use but does not bundle into SOP sandboxes. Antigravity can reuse the execution machine owner's account login and refreshes credentials back to the local store. The seven agent CLIs after Claude Code keep their model and MCP configuration in a private home directory per run, isolating state between sessions.

Illustrative photo: photographer, girl, camera, digital Illustrative image: geralt / Pixabay

The infrastructure model centers on Docker. Both SOP runs and conversation workspaces spin up Docker sandbox containers through the host Docker socket. Default sandboxes reuse installed runtime images, so the platform does not rebuild images after installation unless custom Dockerfiles are introduced. The installer checks for Docker, Compose, Python 3, and curl, offering to install missing components on Linux via apt, dnf, or yum. macOS requires Docker Desktop and Python 3, with Homebrew handling missing dependencies. The service listens on 0.0.0.0:8090 by default, and cloud deployments can restrict access to specific IPs until an administrator account exists.

This sandboxing approach carries a notable security trade-off. Docker socket access grants the service control over the host daemon, which bypasses standard container isolation. Tencent explicitly states that this setup is intended for trusted teams rather than untrusted multi-tenant environments. The platform supports password accounts, project isolation, and member management, but signed-in users can read project data across the instance. Teams can disable public signups by setting AGENTFLOW_ALLOW_SIGNUP=false in the environment file. Remote access requires an HTTPS reverse proxy, and the installer can read public IPs from cloud provider metadata services.

Beyond execution, LoopForge provides visibility into agent behavior and resource consumption. The platform logs node status, tool traces, worker logs, and per-node artifact snapshots in a single dashboard. A Needs review inbox aggregates all human gates waiting for approval, allowing engineers to approve, add context, rerun steps, or send work back upstream. The system also tracks token usage and cost accounting for both SOP runs and conversations at public list prices. Engineering metrics and business-domain reports can be exported as static HTML or SVG files.

The platform exposes an OpenAPI-described HTTP and MCP API, enabling integration with external systems. Teams can connect CodeBuddy to their LoopForge instance over MCP to use platform projects and SOPs directly from a coding assistant. The chat workbench supports @ references for roles, skills, requirements, SOPs, and repositories, and messages queue while a reply runs. Engineers can interrupt and steer conversations mid-reply, pause or resume running SOPs, or take over a workspace entirely.

LoopForge represents a pragmatic step toward productionizing coding agents without requiring teams to build custom orchestration frameworks from scratch. By standardizing the workflow around discrete nodes, explicit human gates, and isolated Docker sandboxes, it addresses the reliability and auditability gaps that typically emerge when agents move from demos to actual codebases. The reliance on Docker socket access remains a constraint for broader deployment, but for internal teams managing their own infrastructure, it provides a structured path to repeatable agent delivery.

Sources

  • github.com: GitHub - Tencent/LoopForge: Self-hosted workflows for coding-agent delivery. Orchestrate CodeBuddy, Codex and
#tencent#loopforge#coding agents#devops#open source#docker
PE
PUBLISHED BY

PNEUMETRON EDITORIAL

Pneumetron's pipeline pairs AI-assisted drafting with human editorial review before publishing — our goal is to make staying informed easier, not to replace real reporting.

Source Material:github.com ↗
Share this article
Share
Stay Informed

Never miss a signal.

Subscribe to the Pneumetron Intelligence Digest.

← Previous
NVIDIA releases Daviator to batch k-points and enable mixed precision in VASP

More from AI Research

View All →
AI ResearchJust now
Illustrative photo: network, connection, pc, web address

NVIDIA releases Daviator to batch k-points and enable mixed precision in VASP

NVIDIA has released Daviator, a GPU kernel library for VASP that batches k-point processing to fix small-system bottlenecks and adds a mixed-precision layer for inference-class hardware, reporting up to 22x speedups over CPU.

BY PNEUMETRON5 MIN READ
Read more
AI Research2h ago
Illustrative photo: engineer, engineering, broadcast, broadcast engineer

Game2Player Scales Visual Agent Training with Co-Developed Game Environments

A new arXiv preprint introduces Game2Player, a framework that iteratively builds and verifies 568 web game environments using agent feedback. Distilling 15.7K trajectories into a 9B model doubles progress on a held-out benchmark and surpasses proprietary baselines on GameWorld.

BY PNEUMETRON5 MIN READ
Read more
AI Research2h ago
Illustrative photo: it, server, server cabinet, network

AWS launches SageMaker model leaderboard; IBM brings time-series models to Confluent

AWS releases a standardized leaderboard for 70+ LLMs inside SageMaker Studio, while IBM and Confluent deploy time-series foundation models directly into streaming pipelines for real-time forecasting and anomaly detection.

BY PNEUMETRON4 MIN READ
Read more
AI Research14h ago
Illustrative photo: glass, reflection, lights, door

AWS Details OAuth Patterns for Bedrock AgentCore to Snowflake Cortex AI

Amazon Web Services published a technical guide detailing three OAuth authentication patterns for routing Bedrock AgentCore AI agents to Snowflake Cortex tools via MCP, addressing enterprise identity and token management requirements.

BY PNEUMETRON6 MIN READ
Read more

Comments

Most Read

01
AI Research·1d ago
Recent AI Research Reveals Gaps in Coding Agents and Advances Scientific Modeling
02
AI Research·1d ago
Four New AI Papers Advance Reasoning, TTS, and Decoding Efficiency
03
AI Research·14h ago
AWS Details OAuth Patterns for Bedrock AgentCore to Snowflake Cortex AI
04
AI Research·14h ago
AWS Agent Registry Introduces Custom Metadata Schemas
05
AI Research·1d ago
NVIDIA Vera Rubin NVL72 debuts in MLPerf Training v6.1 with up to 2x Blackwell Ultra performance
Daily Digest

Get top AI & tech signals delivered to your inbox every morning.

Subscribe →
Follow Signals
X / TWITTERLINKEDININSTAGRAMYOUTUBETELEGRAM

THE PNEUMETRON

The publication of modern technology, AI, and infrastructure.

AI-assisted, human-reviewed.

NEWS CATEGORIES

  • AI RESEARCH
  • TECHNOLOGY
  • WORLD
  • BUSINESS
  • POLITICS
  • SCIENCE
  • HEALTH
  • SPORTS
  • ENTERTAINMENT

AI & RESEARCH

  • AI RESEARCH
  • INFRASTRUCTURE
  • TOOLS
  • DEEP DIVES
  • TECHNICAL DISPATCHES

RESOURCES

  • ABOUT US
  • CONTACT
  • CAREERS
  • ADVERTISE
Pneumetron.
© 2026 Pneumetron. AI-assisted, human-reviewed.
All rights reserved.
ABOUTTOOLSPRIVACYTERMSCONTACT