XNews

Tencent Open-Sources LoopForge for Scheduled Coding-Agent Workflows

Tencent Open-Sources LoopForge for Scheduled Coding-Agent Workflows

Tencent released LoopForge, a self-hosted workflow platform that orchestrates CodeBuddy, Codex and Claude Code through scripts and human approval gates. The team runs five staggered scheduled agents against its own Go repository for research, PR review, CI conflict handling, test coverage and architecture checks, with irreversible actions barred in the first line of prompts and a queue cap of 10 unmerged architecture PRs.

Event version 1.

Key insights

Three-stage workflow model

Work moves from sandboxed conversation with a Main Bot, to a frozen standard operating procedure covering planning, design, coding, script-based deterministic verification, independent review and human sign-off, to scheduled autonomous maintenance.

Five staggered scheduled agents

Agents run daily research at 09:00 with read-only GitHub access and at most three requirement tickets, hourly first-round review, half-hourly CI inspection limited to the automation account's own branches, daily 02:00 test supplementation, and Monday 03:00 architecture checks capped at 200 lines.

Brakes written into prompts

Irreversible operations are placed on the first line of prompts: no merging, no closing, no direct pushes to main, and no touching colleagues' branches. Test agents may only add new test files, and architecture agents stop before starting a sandbox once 10 unmerged architecture PRs exist.

Measured coverage gain

One overnight run raised coverage of a core package from 83.3% to 86.0%. The figures come from one of Tencent's own Go repositories.

Weeks of constraint tuning

The setup took weeks to tune, with most time spent writing constraints. The team states that expecting smooth operation on day one is unrealistic.

Source posts