Optio: Orchestrating AI Coding Agents in Kubernetes from Ticket to PR

Optio is an open-source orchestration system for AI coding agents that manages the complete workflow from task intake to merged pull request without requiring constant human supervision. Built with Fastify, Next.js, BullMQ, and Drizzle on Postgres, it ships with a Helm chart for production deployment.
How It Works
The system follows a three-stage pipeline:
- Intake – Pulls tasks from GitHub Issues, Linear, or manual creation
- Execution – Spins up isolated Kubernetes pods per repo, runs Claude Code or Codex in git worktrees
- PR Monitoring – Watches CI checks, review status, and merge readiness every 30 seconds
The key differentiator is the feedback loop. When CI breaks, Optio feeds the failure back to the agent. When a reviewer requests changes, the comments become the agent's next prompt. It continues until the PR merges or you manually stop it.
Architecture and Features
Optio uses a pod-per-repo architecture with one long-lived Kubernetes pod per repository. Each pod maintains git worktree isolation for multiple concurrent tasks. The system includes:
- Self-healing – Auto-resumes the agent on CI failures, merge conflicts, or reviewer change requests
- Completion automation – Squash-merges the PR and closes the linked issue when all checks pass
- Per-repo configuration – Model selection, prompt templates, container images, concurrency limits, and setup commands tunable per repository
- Code review agent – Automatically launches a review agent as a subtask with separate prompt and model
- Real-time dashboard – Live log streaming, pipeline progress, cost analytics, and cluster health monitoring
Technical Implementation
The architecture consists of three main components:
- Web UI (Next.js on port 3100) with dashboard, task management, and live streaming
- API Server (Fastify) with workers for task queue, PR watcher, health monitoring, and ticket sync
- Kubernetes cluster with repo pods containing multiple worktrees running AI agents
Services include repo pool management, review agent execution, and authentication/secrets handling, all backed by Postgres for tasks, logs, and events.
📖 Read the full source: HN AI Agents
👀 See Also

Reduce AI Coding Session Costs by 90% with Graph-Based Code Indexing
A developer built a local graph database that indexes a codebase using LLM-generated summaries, cutting Claude Code session costs from $6-10 to cents by avoiding redundant file re-reads.

OpenClaw Janitor Skill for Automated System Management and Security Hardening
A developer created a skill that uses Claude Code to SSH into OpenClaw machines and harden configurations, including sandboxing, OS hygiene, and channel security, while maintaining a project folder with audit instructions in CLAUDE.md.

Open-source multi-agent framework extracted from Claude Code leak
A developer extracted the multi-agent orchestration system from Claude Code's leaked source code and rebuilt it as a model-agnostic open-source framework with MIT license. The 8,000-line TypeScript framework includes task scheduling, inter-agent messaging, and built-in tools.

MOOSE-Star: A 7B Model and 108K-Paper Dataset for Scientific Hypothesis Discovery – ICML 2026
MiroMind releases MOOSE-Star on Hugging Face: a 7B model (DeepSeek-R1-Distill-Qwen-7B fine-tune) for scientific hypothesis discovery, alongside the 108K-paper TOMATO-Star dataset. Benchmark shows MS-7B achieves 54.34% inspiration retrieval accuracy, beating GPT-5.4 and approaching Gemini-3 Pro.