Essays

26 essays 6 pillars Evidence base verified 2026-10-19

2026-10-19 Claude Code Memory Usage: How Many Sessions Fit on One Machine Claude Code memory usage, measured: three sessions ran 32 processes and MCP servers outweighed Claude in two. A RAM budget for running sessions in parallel.
2026-10-12 Agent Tool Design Belongs in Your Harness Audit, Not Prompt Tips Agents that sent the same write twice reported success 90% of the time. The fix was one tool parameter. Audit what a tool call means, not just whether it runs.
2026-10-05 Diff Review Can’t See Your Agent Regress. A Standing Suite Can. Per-diff review checks one PR, not whether your coding agent still works. How to run AI agent regression testing as a standing eval suite: triggers, tiers, owners.
2026-09-28 Your Agent’s “Done” Is a Self-Report. Make It a Predicate. Acceptance criteria for AI coding agents fail when done is prose the agent grades itself. Write end-state predicates a checker runs where the agent can’t write.
2026-09-21 Where a Decision Model Belongs: Placing Jev in Your Stack Jev’s launch numbers answer a question you don’t have. The tier it joins already ships at Anthropic, Meta and OpenAI, with the costs published.
2026-09-14 Coordination Is an Architecture Layer, Not a Prompt Instruction When a multi-agent pilot breaks, the postmortem blames the model. The failure data says coordination logic buried in each agent’s prompt is the real defect.
2026-09-07 AGENTS.md vs CLAUDE.md: Swap the Agent, Keep the Harness Claude Code reads CLAUDE.md, not AGENTS.md. Sort your harness into what ports across vendors and what does not, so switching coding agents is a config flip.
2026-09-01 Rank Your Agent Trace Fields by What You Can Never Get Back A conformant agent trace proves the run failed and cannot locate where. Rank the fields you write at call time by what no later analysis can recover.
2026-08-24 The Topology You Can Review Is the Topology You Can Run Eighteen of thirty agents made the same git branch. Default to one agent on coding work, then add an edge only when a named constraint forces it.
2026-08-17 AI Code Review Is a Property of Your Team, Not the Tool Two teams ran the same AI reviewer and reported opposite results. Stop importing a verdict from the discourse and run the diagnostic on your own team.
2026-08-10 Stop Picking Your Coding-Agent Model Off a Leaderboard Your team is about to swap coding-agent models over a two-point leaderboard gap. Anthropic measured six points of movement from infrastructure alone.
2026-08-03 Task Decomposition for AI Coding Agents: Draw the Graph First Across 20,574 agent sessions, the top failure was violating a stated constraint. Draw the task graph before the prompt: nodes, owned files, and contracts.
2026-07-27 Audit Your Agent Harness: The Deterministic Layer Nobody Reviews You have reviewed your CLAUDE.md a dozen times and never opened the file that decides whether the agent’s next rm -rf runs. Start by inventorying what shipped.
2026-07-13 How to Structure CLAUDE.md: It’s a Loading Policy, Not a Document Anthropic closed ‘Claude ignores my CLAUDE.md’ as not-planned, area:model. The fix is structural: route every line to the tier that loads it, not a longer file.
2026-07-06 Loop Engineering Breaks Your Single-Shot Context Playbook A loop was always the agent primitive. The CLAUDE.md budget and JIT retrieval you tuned for one invocation don’t fail louder in a loop — they fail quieter.
2026-06-29 You Can’t Cap What You Can’t Attribute: Per-Task Cost A cost dashboard is a lagging report, not a guardrail. Per-task cost attribution is the schema that makes an agent budget ceiling enforceable before the next call.
2026-06-22 Where Just-in-Time Context Retrieval Silently Breaks JIT context retrieval isn’t free — it’s slower and only as good as your tooling. A ledger of where runtime retrieval breaks and how to design the fallback.
2026-06-15 When One Agent Stops Being Enough: The Isolation Gate The trigger for going multi-agent isn’t parallel speed — it’s context isolation, and the split costs about 15x the tokens. Two gates decide whether it’s worth it.
2026-06-08 Review Capacity Is the Real Ceiling on Your Agents The ceiling on how many AI coding agents your team can run is reviewer-hours, not agent output — and that number is already in your telemetry. Here’s how to size it.
2026-06-01 How to Verify AI Coding Agent Output: A Reviewer’s Framework Generation got cheap and verification didn’t. Six per-task moves for reviewing an agent’s diff — from distrusting its self-report to calibrating by blast radius.
2026-05-25 Tier Your AI Agent’s Production Authority by Task Risk An AI agent deleted a production database in nine seconds. The fix was never a better prompt — it was a permission tier the token should never have crossed.
2026-05-18 CLAUDE.md Instruction Ceiling: Maintained Config, Not a README Your CLAUDE.md doesn’t have a length problem, it has a discipline problem. Budget it against a hard instruction ceiling and maintain it like config, not a README.
2026-05-11 Build vs. Buy Agentic AI: Ownership Is the New Decision Agentic AI cut the cost of building software, not owning it. Build-vs-buy is now a five-year ownership call — priced on the maintenance tail, gated on verification.
2026-05-04 What AI Coding Agents Are Actually Good For (And When to Skip) The internet ranks AI coding agents endlessly and never answers the only question that matters: can you verify the output, and can you undo it if you can’t?
2026-04-27 How to Size Tasks for AI Coding Agents The real constraint is context, not lines of code. Six heuristics for sizing tasks so AI coding agents ship clean code on the first try.
2026-04-20 The Anatomy of a Perfect AI Agent Task Context engineering for AI coding agents — the elements that make a task succeed on the first try, with a fully worked example.