Thoughts & Ideas
On engineering, AI, and building products that matter.
The Context Audit I Run on Every Agent
Trimming an agent’s prompt took 2,146 tokens off every run, and that is the smaller half of the problem. A seven-step audit for the two places context gets wasted, the start of a run and the middle of it, plus the replay rule that covers both.
Asking an Agent Why Production Was Slow at 6:20 pm
An agent my team runs in production answers incident questions in plain English. What made its answers worth trusting: parse the question into a window, fetch a snapshot with baselines, flag anomalies in plain code, reason from symptom to root cause, and cap confidence when data is missing.
A Code Review Agent That Knows Production
Most AI reviewers only read the diff. Ours also reads Sentry errors, Datadog performance data, past incident investigations, learned rules and our best practices, and gets better every week from reviewer feedback.
Seven Things Shipping AI Agents to Production Taught Me
None of them are about which model to pick. Evals before the agent, context as a budget, idempotent tools, loop guards, governed input and output, where people belong once agents write most of the code, and turning your team's taste into checks.
Harness Engineering: Why the Code Around the Model Matters More Than the Model
A model on its own can only generate text. The harness is the code around it that reads files, runs tools, manages memory, and recovers from mistakes. Here is what goes in one, and how Claude Code, OpenClaw, and Pydantic AI build theirs.
How We Got to Modern LLMs
A plain-language walk through every era of language models, from hand-written rules to autonomous AI agents.