ai-engineering

When a supply chain attack lands and the timeline is asking for discipline, the durable fix is one layer down — at the package manager, not at your attention span.

The value of running six frontier models isn't six perspectives — it's six chances to be wrong, which means you can set a much higher bar for what counts as right.

Using cell biology naming not as metaphor but as engineering manual — how mTOR's biology predicted circuit breakers, autophagy, and negative feedback loops before we designed them.

A 96% rejection rate that was actually a 96% false positive rate — how a monitoring blind spot turned a productive overnight batch into apparent failure.

The architect writes the tests. The implementer makes them pass. No prose specs, no circular validation.

I dispatched a 952-line monolithic CLI through my own coding-agent dispatch system to be refactored into seven modules. It worked. Notes on what self-bootstrap reveals about agent harness design.

I scraped 16,181 of Simon Willison's posts and analysed the 395 from 2026. An inflection in November 2025, GLM-5 closing the gap, and why the harness — not the model — is the competitive moat.

I benchmarked four AI coding harnesses on 12 tasks using the same model. The harness barely matters for correctness — it's all about the model.

Your agent system shouldn't stop when the task list is empty. The real bottleneck isn't execution — it's discovering what's worth doing next.

The temptation in agent systems is to make everything a prompt. But most of the work is deterministic — and deterministic work deserves code, not suggestions.

Monitoring systems need consumers before they need features

Most LLM pipelines don't need agents. The ones that do share a specific pattern — the step needs to decide what to do next, not just process what it's given.

Which parts of an AI dev workflow can be wrapped in a CLI, and which can't — learned the hard way by building the wrong thing and measuring it.

Prompt instructions are suggestions. Hooks are constraints. One survives a model swap.

AI commoditises execution. What remains is taste — the 'that's the one' reflex. And the only way to sharpen it is to ship and see what reality says back.

The bottleneck in a world of AI tools isn't crafting the output — it's knowing which output is worth crafting.

AI agents that require you to go looking for their results aren't agents — they're automation with better UX. The loop closes when results arrive, not when you remember to check.

Adding more AI tools doesn't make you faster if you're still the junction between every agent step.

Having great AI delegation tools and not using them isn't a tool problem — it's a pattern recognition problem, and that distinction changes everything.

The execution layer of an AI system is only half the infrastructure — the reporting layer is what determines whether anyone acts on the results.

The gap between AI assistants and AI agents isn't about reasoning capability — it's about whether the thing can survive your laptop closing.

The gap between 'agentic AI' hype and what actually ships in production turns out to be a workflow — and that's a feature, not a failure.

An experiment with AI deliberation revealed something uncomfortable: accumulating confident opinions feels like convergence on truth, but isn't.

I ran an experiment to find where multi-model deliberation adds value. The answer surprised me: it's the structured format, not the model diversity.

The planning/execution split is more useful than any benchmark comparison.

Evaluating LLM systems requires judgment, not statistics. That shifts who's qualified to do it — and where the gap is in most organisations.

When building MCP servers, search should return scannable summaries — not full content. Let the model decide what to read.