The idea

A single coding agent is impressive right up until the task gets big. Then it loses the thread — forgets a decision from twenty steps ago, re-litigates something already settled, drifts off the stack you agreed on.

The fix that keeps working for me isn’t a smarter model. It’s structure. AgentOS is my attempt to give agents an org chart: 22 role-based agents across departments — product, engineering, growth — each with a narrow contract, coordinated by declarative workflows instead of one giant prompt.

You hand it a context pack. It runs a chain of agents to take that from idea to a deployed MVP.

Workflows are files, not prompts

The core decision is that a workflow is a declarative artifact — Markdown and YAML checked into the repo — not a conversation. A product-launch chain or a weekly-growth-loop is a file you can read, diff, and review in a PR.

That matters because it makes the system inspectable. When an agent does something surprising, I don’t re-read a transcript trying to reconstruct what happened. I read the workflow that told it what to do. The behavior lives in version control, next to the code it produces.

The boring problems are the real ones

Here’s what I didn’t expect: almost none of the hard problems are AI problems.

Isolation. When several agents work in parallel, they stomp on each other’s files. The fix is old and unglamorous — git worktrees, one per worker — so parallel agents each get their own checkout and merge cleanly.
Enforcement. Agents will happily reach for a different framework, skip a test, or ship a security hole to close a ticket. So the standard stack and the security non-negotiables are baked into PR gates. The agent can’t merge past them any more than a human could.
Observability. You cannot debug what you cannot see. Everything is traced with OpenTelemetry, because “the agent did something weird” is not a bug report I can act on.

Isolation, enforcement, observability. That’s not an AI wishlist — it’s the same distributed-systems checklist I’d bring to any fleet of workers that share state. Agents are just workers with worse impulse control.

Where it is

AgentOS is on its v2 orchestrator, in active development, with a public beta demo. It’s the project that most directly connects the two halves of what I do: the performance-engineering instinct for systems that stay correct under load, pointed at a fleet of AI agents instead of a fleet of pods.

The model gets the headlines. The infrastructure around it decides whether it actually ships.