You Ship Your Org Chart. Now Your Agents Do.

Organizations which design systems are constrained to produce designs which are copies of the communication structures of these organizations. — Melvin Conway, 1968

Conway’s Law says you ship your org chart. For fifty-seven years that’s been an observation about software engineering. Now it applies to every team using an AI agent. An engineer friend of mine watched this happen to his sales team in real time. They didn’t sign up to be software designers, but the agent quietly promoted them.

Like a lot of engineers, my friend watched his sales colleagues drowning in manual work and thought it would be pretty easy to give them an agent to kill off the drudgery. He gave them all Claude Code with a couple skills for drafting proposals, summarizing calls, and updating the CRM. They started using it and loved it, prompting was faster and better than cutting and pasting.

A couple of the sales reps really got into it and customized their skills, built up their own personal prompt libraries, found other plugins for Claude Code to automate even more apps. They tried to help the rest of the team stay up to date but copying around updated skills and prompts was a PITA, so many of them just stopped updating. So some agents got better and better and others stayed locked in the past.

Meanwhile my friend had given himself a side gig as an agent support engineer. People needed help updating skills. They couldn’t use git. Someone found out they could integrate with Google Workspace but none of them could figure out the auth. Then IT started asking why sales was running Claude Code when clearly that was only approved for the engineering team.

They didn’t ship a “sales agent”. They shipped a “Bob the sales rep agent”, “Mary the sales rep agent”, etc. each with their own quirks. Two questions nobody had asked the VP of Sales: what does our ideal sales process look like, and how would we know if the agent was running it well? They automated fragments of what they had, not what was best. And they had no way to tell the difference.

None of this is an AI problem.

If the team is fragmented, the system is fragmented. The agent didn’t fracture this team, it was the first tool powerful enough to reveal a fracture that was already there.

An agent is software. Every customized prompt is a design decision. Every shared skill is a module. The sales team didn’t realize until too late they were writing software.

When shipping software was slower the fragmentation took longer. Software creation being in the hands of a few naturally constrained it. Product Managers, Project Managers, Engineering Managers and a horde of others all acted as gatekeepers. Not necessarily upholding quality but usually slowing this down enough and funnelling it into a lowest-common-denominator system. Software being slow and expensive to change meant changing organizations around it was easier than changing the software. Now we have the opposite problem.

The agent is the first tool powerful enough to amplify fractures at human pace.

So now all teams adopting agents have to contend with Conway’s Law but they’re doing it without the tools that engineers have. We invented primitives to make Conway’s Law survivable: version control (git), semantic versioning, package managers (npm / pip / cargo), pull requests, code review, continuous integration, changelogs. And we invented primitives to know whether what we shipped was working: tests, regression suites, staging environments, monitoring, telemetry. These are so mundane to a working engineer in 2026 that they feel like oxygen.

You are going to ship your org chart. With agents, you’re just going to do it overnight. The teams that win will be the ones who decide to design the operating model intentionally, rather than letting chaos design it for them.