notes from teams shipping with agents.
essays, postmortems and pattern notes from the steerdev pod. spec-driven, agent-agnostic, evidence-based, written by the engineers actually doing the work.
all writing
showing 6 of 7 posts
The Spec Is the Product: Why Machine-Readable Specs Are the Missing Layer in AI Development
Give three agents the same vague ticket and you get three different implementations, all "correct," none of them what you wanted. The agent did not fail. The input failed. In the agent era, the spec is no longer a memo between humans. It is the primary input to a production system, and it needs to be engineered like one.
Building Fast Is Not Enough. But Neither Is "Alignment." You Need Both.
8090 just published a thoughtful four-part series arguing that speed without alignment builds the wrong thing. They are right. But alignment without speed is just waterfall with better vocabulary. The real question is how you get both at once, and what kind of infrastructure makes that possible.
You Don't Have an ERP Problem. You Have a Lock-In Problem.
NetSuite keeps your books fine. SAP runs the plant. The problem is everything they don't do, the custom module quoted at six figures, the migration quoted at 18 months, the roadmap that learned to stop asking. Those quotes were rational when humans typed every line of code. They aren't anymore. Here is what changed, and what it takes to use the change without shipping slop.
The Great Inversion: When Judgment Becomes the Bottleneck
For forty years, writing the code was the bottleneck. With capable agents in the loop, that share is shrinking, and a different cost is rising in its place. Here is what the new shape looks like on a real codebase, and what it does to the engineers around it.
Your Team Doesn't Have an Agent Problem. It Has a Coordination Problem.
Monday morning, five engineers each spin up two or three coding agents. By Wednesday, three pull requests collide, two agents have rewritten the same module, and nobody can trace which spec drove what. The agents aren't broken. The team coordination layer is missing.
Agent-Agnostic Is the Only Sane Default (and Here's the Data to Prove It)
In the last eight weeks the agent landscape has shifted three times. Opus 4.7 became the default across four platforms, Cognition shipped SWE-1.6 for free, OpenAI launched GPT-5.5, and Warp added three open-weight models in a single release. If your team standardized on one agent two months ago, your choice is already outdated.