Designing an Operating System for Human-Agent Teams

I run part of my engineering team with AI agents.

Not as a cute autocomplete.
Not as a talking sidekick for developers.

I mean full-fledged autonomous agents taking on well-defined roles in our engineering workflow.

I didn’t start out with a plan to do this. As a product leader without a traditional software development background, my curiosity got the best of me. I wanted to explore how far we could push AI to become an active contributor, not just another tool to write code, but a partner to help reimagine how we build software itself.

Here’s what my engineering workflow looks like today:

1. I — Product/Human Lead: I define the problem, set priorities and acceptance criteria, and determine what “good” should look like.
2. Claude Code — Lead Engineer: This AI agent does the heavy lifting, taking on much of the initial implementation work directly in the codebase.
3. GitHub — System of Record: Issues, branches, pull requests, version control, and CI all operate here.
4. Codex — Staff Engineer/Reviewer: Another AI agent that focuses on reviewing architecture and code, surfacing risks, and collaborating on implementation decisions.
5. QA Agent: It tests the product against acceptance criteria, hunts regressions, and highlights areas where things don’t meet the bar.
6. Security Agent: This agent reviews authentication, ensures data remains safe, and identifies vulnerabilities.
7. Me — Human Approval: I make the final call: what gets accepted, what needs more work, and what actually ships.

What fascinates me most isn’t just the models. It’s the operating model around the models.

A Lesson from the Field

Microsoft’s 2026 Work Trend Index highlighted a compelling point: at an organizational level, the biggest drivers of productive AI use are not just the technology itself. The study found that culture, management practices, and thoughtful workflows had roughly twice the impact on outcomes as individual behavior alone. Their most advanced teams also stood out for something else—they were deliberate. They made intentional decisions about what work to offload, clearly defined human-agent handoffs, and established standards for quality.

But there’s also a cautionary note. A 2025 study by METR showed that experienced developers working in familiar codebases took 19% longer to complete tasks using early AI tools. Why? It turns out that adopting AI without tuning the process can slow teams down, not speed them up.

The takeaway?

AI ≠ faster engineering.

The more interesting insight is:

The Workflow Matters

To succeed, we need to ask (and answer):

• Who writes the code?
• Who reviews it?
• Who tests it?
• What permissions does each agent get?
• Where does human judgment become mandatory?
• Who ultimately owns the result?

Humans + Specialized Agents

In the years ahead, I believe more engineering conversations will turn to questions like these. The future won’t just be engineering teams using AI.

It will be teams of humans and highly specialized agents, working together with clearly defined roles, processes, and handoffs.

I’m still experimenting with what this looks like in practice. But it’s becoming clear to me that one of the essential skills for leaders in the coming decade will be designing and managing what I call human-agent operating systems.