Blog
Graph Engineering: When an Agent Loop Should Be a Graph
Graph engineering moves agent control flow from inference time to authoring time. When a loop should become a graph, and what each pattern actually costs.
Your Agent Swarm Is Probably Just Expensive Ensembling
A multi-agent swarm can beat one model on decisions, but debate often loses to self-consistency at equal compute. Four conditions under which the swarm actually earns its cost.
The Verification Burden Your AI Product Ignores
Accuracy is a lagging indicator. A controlled study found AI made expert developers 19% slower, because the work moved into reviewing. Here is the number to optimize instead.
The Shape of a Coding Agent: 544 Lines That Matter
I read Tau, a 225-file terminal coding agent, as a knowledge graph. The whole coding agent architecture turns on one idea. The event stream is the contract.
Document Parsing for AI Agents
How to choose between OCR, structured parsers, VLMs, and single-shot parsing.
10 Lessons From Production AI Agents
Ten practical lessons from tracing, evals, prompt versioning, routing, and security.
Security for Agents That Actually Do Things
Agent security requires scoped tools, sandboxing, policy checks, and audit trails.
What Self-Improving Agents Actually Mean
Self-improving agents need traces, evals, safe patches, and human approval loops.
Model Routing for Practitioners
Route simple tasks to cheaper models and reserve frontier models for hard calls.
Treating Prompts Like Code
Version prompts, tool definitions, and eval gates so production changes are auditable.