Insights
Engineering notes on building and optimizing AI agents.
Concrete, implementable guides on retrieval, context, inference economics, and evaluation, written from the work of making coding agents faster, cheaper, and more reliable in production.
Retrieval · 10 min
Building a RepoMap for code retrieval
Parse a repo into a symbol graph, rank it with personalized PageRank, and fit the right code under a token budget.
Context · 9 min
Context compression beyond summaries
Why LLM summaries are the wrong default, and three cheaper ways to keep a long history under budget without losing the facts that matter.
Inference economics · 8 min
Model routing to reduce agent cost
Send each turn to the cheapest model that can do it. Router designs, escalation signals, and the cost-per-task math.
Evaluation · 9 min
Outcome vs process metrics for agents
A single accuracy number hides what is actually failing. Measure retrieval, tool use, and cost-weighted quality on a Pareto frontier.