Skip to content
Tencent Open-Sources LoopForge: AI Agent Orchestration for Team Development Workflows

Tencent Open-Sources LoopForge: AI Agent Orchestration for Team Development Workflows

Tencent's LoopForge platform turns personal AI coding agents into configurable team roles, defines executable SOPs with automated feedback loops, and runs 24/7 in isolated cloud sandboxes to make software delivery repeatable and auditable.

SmartApe AI
SmartApe AI
SmartApe AI

From Personal Agents to Team Roles

LoopForge lets each team member codify their expertise as a reusable role. A role consists of four parts: Identity (name, responsibility, executor such as CodeBuddy, Codex, or Claude Code, and model), Task (prompt specifying work sequence, upstream inputs, and constraints), Artifacts & Gates (required output files and pre-submission checks), and Skills & Tools (team standards, Skills, and MCP tools). For example, a Backend Engineer role can be configured with a specific executor, a prompt that reads requirements and design docs, mandatory unit-test artifacts, and access to the team's linting Skill. Roles are versioned; any change creates a new version that can be rolled back instantly.

Shared Context: One Main Bot for the Whole Team

Instead of scattered chat windows, product managers, developers, testers, and leads join a single project and converse with the same Main Bot . The Main Bot acts as a team assistant: it handles trivial tasks directly (small bug fixes, draft designs, code searches) and delegates specialized work to the configured roles (e.g., asking the Backend Engineer to modify an API, the Test Engineer to add cases). All participants share one conversation history — previous messages, role outputs, and workspace changes — so anyone joining later sees the full context and continues seamlessly.

Codifying Workflows as SOPs with Three Node Types

Teams define a Standard Operating Procedure (SOP) by chaining roles in order. Each step becomes a node of one of three types:

Role nodes — judgment-intensive work like code changes or reviews, executed by the assigned AI role.

Script nodes — deterministic tasks such as builds or test runs; success is determined by exit codes, not AI claims.

Human gates — decision points requiring a person's approval (e.g., final sign-off by the lead).

If a node fails, the SOP routes the work back to a previous node with the failure reason, creating a Loop . A retry limit prevents infinite loops; after the configured number of retries the flow pauses for human intervention. The article illustrates a backend-feature SOP: Requirement Analyst → Backend Engineer → Code Reviewer → Test Engineer → Lead Acceptance, with automatic rollback from Reviewer to Engineer on failure.

Visual Execution Trace and Full Audit Trail

Every SOP run produces a visual timeline: green for passed, red for failed, amber for awaiting confirmation. Clicking any node reveals code diffs, tool calls, failure causes, and retry history. All items needing human action appear in a "Pending My Confirmation" list. The complete record — original requirement, SOP version, participating roles, rollback reasons, approver, and final PR — is retained for audit.

Continuous Improvement via Intelligent Evaluation

After a run, an "Intelligent Evaluation" analyzes the execution trace across dimensions: process health, latency and wait times, tool usage, token efficiency, failure recovery, and delivery completeness. In the example, the evaluator found that a missing permission check surfaced only at code review because neither the requirement template nor the review checklist included it. It proposed two concrete updates: (1) Requirement Analyst must specify data permissions for query/export features; (2) Code Reviewer must add privilege-escalation checks and permission tests. The AI only suggests; the lead decides which changes to adopt and updates the role definitions or SOP structure accordingly. Subsequent runs then enforce these checks by default.

Metrics-Driven Validation

A statistics dashboard tracks first-pass rate, rework rate, node duration, human interventions, and token cost. Different SOP versions can be compared on success rate, duration, and cost, enabling data-backed retention or rollback of process changes.

Cloud-Native 7×24 Operation

LoopForge moves the execution environment to the team's own servers or cloud VMs. Each SOP launch spins up an isolated Docker sandbox that clones the repo, prepares dependencies, and creates a branch. All development, review, test, and PR steps run inside this sandbox. When the flow waits for human approval, the sandbox can be released while the full task state is persisted; on resumption a fresh sandbox picks up exactly where it left off. Multiple requirements run in parallel without interference.

Scheduled Automation

A cron-like scheduler periodically scans the requirement pool and dispatches pending items to the appropriate SOP, skipping already-processed or closed items and respecting a configurable concurrency cap. Scheduled jobs can also run prompts that summarize recent runs — failures, stuck flows, pending approvals — and push the digest to the team chat.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

AI agentsteam collaborationcontinuous improvementCodexClaude CodeDocker sandboxLoopForgeSOP workflow
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.