Writing.

Engineering write-ups on how I give coding agents memory, split work between models, and carry state across context limits. Written in July 2026.

A memory OS for agents

Four memory layers, 2.9 million embedded chunks, and an auditor that marks a stored fact stale when its source changes.

Conductor economics

When the strongest model should plan and a cheaper one should type: the dispatch contract, the validation gate, and the arithmetic.

Self-compaction

Checkpointing an agent’s working state before its context fills, and letting it choose when to compact, with public source code.

The engineering behind the essays

See the engineering