omni: an agent orchestrator for my GitHub workflow

Overview

In March 2026 I let an autonomous agent system run my GitHub workflow overnight for the first time.

By the time I woke up, it had completed 686 cycles, opened and merged 11 pull requests, and self-healed 8 stalls without me touching the keyboard.

That system is omni, a CLI orchestrator I use as a personal operating system for AI agents.

What omni does

omni is a 3-layer agent orchestrator:

1. Master layer, decides which workspace or project needs attention next.
2. Workspace layer, understands a specific codebase, open issues, and local tools.
3. Project layer, runs Claude Code agents against a single GitHub issue until it is implemented, tested, and ready to merge.

Instead of treating AI as a one-off assistant, omni treats agents as long-running workers with a backlog:

• It reads GitHub issues and converts them into concrete tasks.
• It spins up short-lived Claude Code agents to implement changes in local repos.
• It commits and pushes branches, opens pull requests, and responds to CI feedback.
• It logs failures and friction points back into a vault, so the system can improve over time.

The goal isn't to replace engineers. The goal is to give a senior engineer a swarm of reliable, scriptable collaborators.

Why I built it this way

Most "AI coding" setups stop at:

"Open the editor, ask the model to change a file, copy/paste a diff."

That works for small changes, but it breaks down when you want:

• A queue of GitHub issues that get implemented automatically.
• Agents that can run commands, read lints, and keep state between steps.
• A feedback loop where every failure becomes a future improvement.

omni is my answer to that. It:

• Treats GitHub issues as the source of truth.
• Uses Claude Code agents as workers, not as a UI gimmick.
• Stores context and run history in a vault, so agents can learn from previous attempts.

The result is an OS-like layer that sits between my repos, GitHub, and Claude Code.

The first overnight run

The first real test for omni was simple: "Run all night and try to close as many issues as you can."

Over one night it:

• Ran 686 orchestration cycles across multiple projects.
• Opened and merged 11 pull requests.
• Hit 8 stalls (e.g. flaky tests, ambiguous specs) and recovered from them without manual intervention.

The interesting part wasn't the raw numbers, it was the failure modes:

• Places where the agent got stuck in tool loops.
• Missing lint rules or tests that needed to exist.
• Ambiguous GitHub issues that a human would have rewritten without thinking.

Every one of those became a new constraint for the system. The next night, omni knew a bit more about what "done" looks like.

How it fits into my work

omni isn't a SaaS product (yet). It's a tool I run locally as part of a bigger stack:

• context-vault provides a persistent memory layer for agents.
• This site (felixhellstrom.com) is an agent-native static site, hand-edited HTML deployed via Cloudflare Pages.
• omni runs on top of that environment, spinning up agents that can safely edit repos like this one.

If you work with:

• A lot of small issues across multiple repos.
• Repetitive refactors or migrations.
• Documentation and content updates that agents can do faster than humans.

then omni is the kind of system I like to design and implement.

Work with me

If you're interested in building:

• Agent orchestration systems like omni,
• MCP-based tools like context-vault, or
• Production Webflow / HubSpot CMS setups that integrate with AI workflows,

you can get in touch here. I work full-time as Technical Lead at Stormfors and take on a small number of projects through my independent practice.