Architecture Is What Agents Can't Refactor
Agents made code structure cheap to change. What's left of architecture is what a revert can't undo and what needs someone else's agreement, so that's where platform teams should spend their design time.
★topic · 33 essays
Platform engineering is the work of taking decisions every team would otherwise make differently and turning them into shared defaults, templates, and guardrails. Done well, the right way is also the easy way. I've built two greenfield cloud platforms in financial services. Most of what I believe about this work comes from the choices that aged well and the ones I'd redo.
A few positions run through these essays. Platforms only work when defaults are inherited, not chosen. Removing direct production access is a culture change before it's a tooling change. Complexity is unavoidable in a large organization, and the drag it creates is something you can design out.
The newer essays ask what changes when the platform gets a second user. Developers now arrive with an AI agent beside them, and the agent can't work around the ambiguity people absorb every day. That makes agent readiness a useful test of platform maturity. It also moves work onto the platform: trusted context, an interface that accepts intent, an identity for the agent, and a way to stop it.
Abundant generation makes constraints more valuable. An agent can produce hundreds of technically valid solutions, and none of them tells it which one fits your security model, cost expectations, or support model. So the paved road matters more once agents arrive. It's the boundary between what an agent can create and what the organization is prepared to operate. Every default, error message, and exception also teaches engineers and agents how work gets done, whether or not anyone chose the lesson.
Start with the essays below. The Agent-Ready Platform series takes the second half of this argument in order.
start here · 9 essays
The pre-agent thesis: good platforms convert unavoidable complexity into defaults and automation.
Two greenfield cloud builds in financial services: the decisions that aged well, and the ones I'd redo.
Removing production access is a culture problem, and culture problems do not yield to Terraform.
Names the new user: the platform now serves the developer and the developer's agent.
The first capability: trusted context, provided by the platform instead of gathered by hand.
The interface moves up: developers state intent, and the golden path becomes the compiler target.
Defaults, errors, and exceptions show engineers and agents what normal work looks like.
An agent acting on your authority needs its own identity and a stop the platform can prove.
Follows a saved hour across the delivery journey, so platform teams can see what they remove.
go deeper
Seven essays on platform engineering after the agent arrives.
Six checks of whether an agent could deploy without asking anyone.
What the paved road settles, and what it never fixed.
The delivery side of the same argument.
the full list · 33 essays
Agents made code structure cheap to change. What's left of architecture is what a revert can't undo and what needs someone else's agreement, so that's where platform teams should spend their design time.
★Writing my transition plan showed me what our metrics could not establish: what people had learned, how their judgment had developed, and what my leadership had contributed.
★An agent that saves you money threatens somebody else's revenue. Amazon, Walmart, the banks, and the card networks are deciding who gets to influence the customer's decision, and better authentication settles only part of that argument.
★AI is making some old flaws easier to exploit. Black October exposed the harder problem: delivering a safe fix when both delay and the newest release can put you at risk.
★A change can pass release checks after an agent has exceeded its authority. Platforms need controls during execution, with evidence that those controls hold even when the model follows a malicious instruction.
★In May I named the argument underneath a year of writing. Five months later, I'd keep the argument and narrow several claims. Here is what changed, what survived, and what I still can't prove.
★Building a portable software-development workflow across Claude Code, Codex, Goose, Pi and OpenCode.
★An agent acting with your credentials can leave a record of your authority without a record of its involvement. Platform teams need to preserve who acted, under whose authority, and how to stop that access.
★Every default, error message, and exception communicates how work gets done. The question is whether that's the lesson you intended.
★Omarchy installed on a drawer laptop in two minutes and everything worked. The reason is a platform lesson: choose the defaults, keep the escape hatches, and make the system legible to the agent.
★Kiro's frontier engineering principles describe how to work well with agents. Making that practice dependable across a team means treating the shared workflow as a software release, with a version, a scope, an owner, and tests for when it should stop.
★An AI tool can save a developer an hour while the delivery journey consumes most of the saving. The developer experience ledger follows the work across that boundary, so platform teams can tell what they removed and what they merely moved.
★Sixteen months of building and writing taught me that AI makes implementation cheaper while raising the value of clear intent, trustworthy platforms, operational understanding, and human judgment.
★The obvious version of AI developer experience is a chatbot in the portal. The better version diagnoses the failure, repairs what the platform owns, and reruns to green before the developer opens the log.
★A gate proves a standard ran. It does not prove anyone can operate the result at 3 a.m. Operability has to become part of the spec, and the evidence has to survive the merge.
★The golden path made the right way easier to follow. It still asks developers to learn the road. With AI, intent can become the interface while the platform keeps the standards, the controls, and the gate.
★We treat enterprise context as something developers gather before they can use AI well. Platform teams should treat trusted context as something the platform provides.
★We built internal platforms for developers. Those developers now arrive with an AI agent beside them, and how well your platform serves that second user is becoming the real maturity test.
★Complexity is inevitable in modern engineering. Drag is optional. The work of leadership is turning repeated friction into platforms, defaults, and systems that help teams move faster with trust intact.
★A year of writing, one argument, and a working theory of where software is going
★Code used to be the durable asset. In an agentic SDLC, that changes. Code becomes the regeneratable output of a system that runs on something more important: a clear, versioned, reviewable specification. That shift changes what engineering organizations invest in, how they govern delivery, who they hire, and what they actually ship.
★Jasmine Sun argues AI politics has a new meta, and the warning shots have started. Reading her piece as an engineering leader, here is what the narrative failure looks like from inside a large team, why sociopolitical alignment is our job, and what each of us owes our own career in a market this fast.
After two greenfield cloud builds in financial services, these are the decisions that aged well, the ones I would redo, and why the small choices in year one decide whether you have a platform or a pile in year five.
★The strategy posts say AI software development is a system. Here is the working loop I run inside that system: a refined specification, a layer of standards, and a coordinated set of specialists doing the work.
★After thirteen months of daily Claude Code use, I stopped treating AI coding as a prompt discipline problem and started treating it like an engineering system: configurable, layered, observable, and built to learn.
Everyone sells pipeline-first delivery as a best practice. Remove access. Route through automation. Enforce consistency. What the slide deck leaves out is the part that actually determines whether this works.
★Back in January I wrote about OpenClaw as a concept. That post was the theory. This one is the implementation. One month running a dedicated EC2 instance, a name, and an agent I genuinely rely on.
Most enterprise conversations about GenAI are arguments about assumptions nobody has questioned. Here is what stays when you strip everything else away.
★Anthropic has shipped more meaningful product features in the last few weeks than most teams ship in a quarter. A theory, and what it tells us about what AI-assisted development actually unlocks when a team uses their own product to build it.
The companies that built the modern internet went bankrupt doing it. The companies building AI infrastructure may follow the same path. That is not a warning. It is how transformative technology actually works.
A humorous confession about subscribing to nine different AI development tools in my quest to find the best tool(s).
AI-powered interfaces are evolving beyond traditional graphical UIs toward intent-based interactions, where users describe desired outcomes rather than navigating through menus and clicks.
A hands-on look at how AWS Kiro transforms AI-assisted development from chaotic vibes into structured, spec-driven engineering.