New in the Lab: Ghostctl, a command-line tool for managing Ghost instances at scale.
Field Notes
Writing on architecture, AI, and transformation, and the judgment required when systems stop behaving as expected.
Running ten agents can still be a single-player experience. The next step is shared work that people can join, challenge and take over, with permissions and evidence that survive the original conversation.
Trustworthy delegation begins when intent, context, authority, guardrails, and evidence become properties of the system rather than assumptions in a prompt.
A trace can show what an agent changed and who approved it. It cannot prove that anyone exercised judgment. When nobody reads every line, control depends on evidence, authority, and the power to stop.
An agent cannot respect a decision it cannot find, scope, and test. Architecture must become usable context before it can become an effective control.
As agents drive down the cost of a first draft, the bottleneck moves to deciding what should exist and producing enough evidence to trust the change.
The platform capability most tied to a good developer experience is not features. It is clear feedback on what happened. Logs, diagnostics, repair hints: that is the product, not the support around it.
Shift left moved the work earlier and dumped it on developers. Shift down moves the complexity into the platform, so they never have to carry it at all.
Loop engineering pays off when a platform's contracts, tests, and policy checks decide whether an agent's loop is actually done.
For thirty years we built developer platforms for developers. The next ones are built for the agents developers delegate to.
A golden path built for humans is a recommendation. A golden path built for a machine has to be an interface.
A platform team that rules by doctrine gets compliance. A platform team that runs like a product gets adoption. Only one of those is worth anything.
A platform team is not a cost centre. It changes the marginal cost of safe change, and that is an economic argument, not a tooling one.
AI will make some teams smaller. It will not make the work lighter.
The next phase of AI-assisted engineering will not be more magical. It will be more explicit.
Agentic engineering is the discipline of making software work explicit enough that machines can safely help with it.
Platform engineering value comes from reducing organizational friction, lowering the marginal cost of safe change, and making delivery evidence visible.
Human oversight in AI systems must be designed as a control capability, not treated as a reassuring box at the end of a workflow.
Event contracts used to be guidance. With agents on the other end, they're the thing the system actually runs on.
The real architectural decision isn't how fast you process data. It's what you choose to treat as a fact, and whether the machines around it can reason about those facts without asking permission.
Modern data platforms are becoming operating systems for decisions, automation, governance, and the flow of organizational work.
AI-assisted development rewards engineering clarity because faster execution amplifies the quality, ambiguity, and governance of the system around it.
Alfred shows why agentic engineering starts before code: tickets need enough context, constraints, and evidence for machines to help safely.
Technical value assurance closes the gap between delivery and business impact by making value, risk, evidence, and ownership visible.
AI-assisted PR review only works when agents can bring issue context, implementation evidence, and review intent into the same engineering surface.
A Terraform-backed Bluesky handle turns social identity into a small infrastructure problem: explicit DNS, owned naming, and portable presence.