Writing
Notes on orchestration, infrastructure, and the methods I want to reuse.
14 posts
The collaboration harness
Two names already exist for this stack: one wraps a model so it can act, one governs several of those wrappers. A live call between a human and recruitable agents needs a third.
A picture is worth a thousand tokens
Making diagrams an agent's default output format took more than a skill description: trigger stacks, context budgets, a rendering failure, and a last line of defense.
When the repeating unit is the product
For platforms of many near-identical units, the product is the scaffold: discovery → analysis → generation → validation, with CLI tools collapsing context and scripts judging while agents propose.
How a diagram-to-infrastructure tool turned into a research habit
Terragram kept losing the thread on long sessions. Fixing that meant studying how models use context, and that research became the reusable asset.
Orchestration engineering: systems that spawn engines
After prompts and context, what compounds is meta-engineering: reusable orchestration that generates domain engines with roles, gates, and encoded expertise.
Let the scripts judge and the agents fix
The most reliable pattern I keep reaching for: deterministic checks decide, and agents only ever propose.
Nobody should both write and merge
Separating edit authority from merge authority is the one rule that makes agent-authored code safe to ship.
One service layer, many surfaces
A browser, a Claude agent, and a background worker should all go through the same code. Anything else drifts.
Compress for agents, not for humans
Context windows are attention budgets. Optimize value-per-token with offline distillation and structured handoffs, not prettier prose for people.
Why I build orchestration systems locally first
Reproducibility before scale: run agent loops on your own machine before they touch anything shared.
Studying how hidden information breaks
You can’t reason about the robustness of a watermark or a steganographic scheme until you’ve tried, carefully, to break it.
Agents deserve a real context format
AGENTS.md and SKILL.md solved pieces of agent context. BRIEF.md is my attempt at the missing discovery-and-handoff piece.
Analogical reasoning as an engineering tool
Borrowing the shape of a solved problem speeds up the next one, as long as I stop where the domains diverge.
Dual channels and rule layers beat a longer prompt
When sessions grow, better prompts help less than splitting silent generation from review, and pushing recurring guidance into a cheap rule hierarchy so the working window stays small.