Insights

Field notes from running AI agents in production: governance, coordination, and organizational design. Written by our team, human and AI, and accountable to a human either way.

The longer arcs live on Substack: Jobs, Skills, and Company of the Future →

Empirical Synthesis

Six Weeks Running a Five-Agent AI Swarm: What We Learned the Hard Way

We ran a five-agent production swarm for six weeks and built a governance framework because we had to. Here's the unvarnished account of what coordination failure actually looks like — and how we fixed it.

13 min readRead
Empirical Synthesis

A2A, MCP, ACP: Making Sense of the Agent Protocol Wars

Four major agent communication protocols dropped in the last eight months. Each claims to solve a different problem. Here's how to think about what you actually need to implement.

8 min readRead
Accountability Infrastructure

Agent Swarm Architecture: The Pattern Nobody Teaches You

Most multi-agent systems fail not because the agents are bad, but because nobody designed how they'd coordinate. After running a 5-agent production swarm for six weeks, we found three structural failures that kill almost every swarm in the first month. Here's what to build instead.

11 min readRead
Accountability Infrastructure

Who Is Liable When Your AI Agent Makes a Mistake?

When an AI agent takes an action that causes harm, the accountability question doesn't answer itself. Here's the framework for thinking through who owns what—before a regulator or plaintiff does it for you.

9 min readRead
Accountability Infrastructure

Your AI Governance Problem Is an Org Design Problem

Most enterprise AI governance programs fail within 12 months. Not because the policies were wrong — because no one owned enforcement. The policy exists. The owner doesn't. Here's what the organizations that get this right do differently.

9 min readRead
Organizational Design

CellOS and the OWASP Agentic AI Top 10: Two Layers of the Same Stack

OWASP tells you what can go wrong. CellOS is the answer to how you structure your agents so those things become structurally harder to do. They're not competing — they're two layers of the same stack.

5 min readRead
Organizational Design

Five Roles Every AI Team Needs (That Nobody Talks About)

Most AI teams are built for the demo, not for production. Here's the org design nobody ships with their agents.

8 min readRead
Accountability Infrastructure

How to Write a SOUL.md for Your AI Agent

SOUL.md is the governance document your AI agent needs before it ships. Here's the structure, what to include, and the five gaps that kill most attempts.

9 min readRead
Accountability Infrastructure

When the Proxy Becomes the Principal

An AI agent ran 315 audit reports with zero behavioral changes. The audit pipeline had become a content pipeline. This is Goodhart's Law at the organizational level — and it reveals why most agent deployments fail not from bad code, but from missing specification authority.

10 min readRead
Coordination

Why Multi-Agent Systems Fail: The Coordination Breakdown Pattern

Three months into running a production multi-agent swarm, the failures we anticipated (model errors, hallucinations, API timeouts) weren't the hard part. The hard part was coordination. Here's the pattern we keep seeing—and how to prevent it.

7 min readRead
Empirical Synthesis

We Gave an AI 100 Experiments and a Gaming GPU. It Rewrote Our Assumptions About Autonomous Research.

An autonomous AI agent ran 100 ML experiments overnight on a consumer GPU and improved its target metric by 57%. The model it trained is irrelevant. What matters is what the experiment revealed about how autonomous agents actually do research.

12 min readRead