How a 24/7 AI Community Admin Handles PRs at 2 AM with AgentTeams

Problem

Open‑source maintainers of the LoongSuite suite faced a flood of pull‑requests (PRs) and issues. Responses taking longer than 48 hours caused contributor churn, while CI failures, complex CI environments, issue triage, and cross‑time‑zone coverage added operational pressure.

Initial attempts and pain points

Agent lifecycle management – custom glue code was required for scheduling, timeout handling, context persistence and failure recovery.

Credential management – GitHub PATs and LLM API keys were scattered in scripts and environment variables, exposing clear‑text secrets.

Lack of collaboration infrastructure – agents communicated via files or DB polling; no unified messaging layer existed.

Zero observability – failures could only be traced through logs; no end‑to‑end tracing was available.

Why AgentTeams

AgentTeams provides a multi‑agent governance platform built on the Matrix protocol. It offers transparent, auditable communication, zero‑trust credential isolation via an AI gateway, and native Team Leader‑Worker hierarchy. These capabilities let developers focus on “what the digital employee should do” rather than on runtime plumbing.

Architecture and implementation

The core component is the github-manager Worker Agent, running on the QwenPaw runtime. Its “brain” is a large language model (LLM); the platform supplies scheduling, tool invocation, messaging and credential management.

Key capabilities:

Five cron jobs provide continuous 7×24 inspection.

Eleven modular Skill directories (each with a SKILL.md and scripts) implement reusable procedures such as PR review, issue classification, CI auto‑fix and daily summary.

The code engine ( Qoder CLI) generates specifications and applies code modifications; a workflow agent orchestrates calls to the code agent.

Operational results (3 weeks)

108 PRs and 48 issues processed.

24 direct code fixes pushed to contributor forks.

First‑response time for new contributors dropped from days to under 1 hour.

CI failures were auto‑fixed: PR #229 missing an Apache 2.0 header was corrected, committed and a comment posted with the new SHA.

Review comments were anchored to specific lines and evaluated on correctness, performance, tests and compatibility.

Maintainers spent less time on repetitive tasks and more on decisions requiring human judgment.

Key lessons and fixes

Timeout configuration – the default timeout_seconds=120 was too short for PR scans (3‑5 min). All tasks were aligned to 900 seconds and the audit task timeout was increased, preventing silent kills.

Approval stamp pitfall – bulk approving based only on CI green led to nine low‑quality approvals that were later dismissed. The Skill now forbids skipping the full review flow.

Self‑identification – the agent mistakenly treated its own /approve comments as external feedback, causing feedback loops. Adding a dynamic gh api user check filters out self‑generated events.

Technical quirks – cron shell escaping required backslashes for arguments such as --text $(date +%Y-%m-%d). APPROVE idempotency was ensured by checking the current HEAD SHA via gh pr view --json reviews (the GraphQL commit_id field is null, so the REST API is needed). Event‑log performance was improved by batching Python processes instead of spawning one per event.

Future directions

Specialized workers (PR Review, Issue, CI) coordinated by a Team Leader agent.

Full Issue→PR pipelines: an Issue Agent classifies, a Spec Agent generates a design, a Code Agent implements, and a Review Agent validates.

Cross‑repo coordination via a Manager agent that detects breaking changes in one repository and notifies workers in dependent repos.

Adopting the Team Leader‑Worker architecture to enable parallel processing of large PRs and end‑to‑end workflow handoffs.

When to adopt a digital employee

Suitable scenarios feature standardized, rule‑based processes that can be encoded as Skills, require 24/7 coverage, and involve multi‑party collaboration (human‑Agent or Agent‑Agent).

Unsuitable scenarios need high creativity, fuzzy judgment or extremely low tolerance for error.

Conclusion

The 24/7 AI digital employee transformed the LoongSuite community experience: a PR submitted at 2 AM received detailed, line‑anchored feedback within an hour, freeing maintainers to focus on strategic decisions. AgentTeams proved to be a cloud‑native multi‑agent collaboration platform that supplies runtime, scheduling, security, observability and governance needed for production‑grade AI agents.