What I mean by an agent harness
Cursor and Codex are the cockpits. The durable harness is the portable, tested system of skills, policy, memory, gates and feedback that survives between them.
Read and shareChop Wood Carry Water
Durable Agent Harness
I’ve been building a working relationship with coding agents — now Codex-first, with portable skills, memory, signed patterns, human gates, and proof close to the work. Models and cockpits change; the practice should survive them. This site is the public notebook.
Agents: start here — pack, Priority 0 (P0) rules, and the 60-minute path.
Start in 60 minutes
Then pick another door
You don’t have to read the whole notebook in one go.
Do
Playbook
Fifteen steps if I were starting from scratch tomorrow.
Move
What Codex changed
Portable skills, layered guidance, worktrees, self-evals — and the parts that did not transfer.
Copy
Starter kit
Thin templates — skills, rules, patterns, lanes, clocks.
Learn
Kaizen → Toyota Way Practices
Improve on purpose; name the practice that matches the pain.
Run
Preloop + OpenRouter
Runtime policy under human gates; route models when it helps.
Ops
Channels & lean
Messaging, ingest, token rent, practice surfaces.
Browse
Explore
Timeline, stack, capability, patterns, glossary, research.
Track
Measure
Clocks and rituals — no slide-deck multipliers.
Read
Insights
What broke, what locked, what earned a place on disk.
The cockpit changed. The work did not.
What the interim earned
What did not transfer automatically
Build your own working relationship
Anti-patterns
Kaizen → Hansei → Toyota Way Practices
MemPalace for memories
Do
Not
Toyota Way Practices — member practices
Runtime layers beyond the cockpit
PR / commit checks
Do
Not
Channels, media & token lean
Messaging, media ingest, token rent, and practice surfaces — one operating system, same human gates.
Do
Not
Do
Not
Anti-patterns
Explore
One panel at a time — timeline, stack, capability, patterns, glossary, research.
Case studies
Measurement kit
Four clocks
Humbling baseline: METR’s early-2025 RCT found experienced open-source developers were about 19% slower on familiar issues when AI tools were allowed — while still believing they were faster. That does not kill the case for agents; it kills lazy multipliers.
Comparison bands
Weekly ritual
Starter kit
Insights
When answers are cheap, judgement matters more
AI can produce answers in seconds. Education should bring critical thinking back to the centre, with the Trivium offering a useful historical lens.
Read and shareA quick check is still a model turn
Repeated fast diagnostics can burn a large cached context, confuse unlike usage ledgers, and turn a read-only review into an unintended change.
Read and shareThe cockpit changed. The work did not.
Moving from Cursor to Codex separated the durable working relationship from its host: portable skills, native policy, worktree discipline, self-evals, and proof.
Read and shareThe sensor must cost less than the work
Included usage is not free. Watcher agents that monitor GitHub or spend are the anti-pattern. Timers, caches, and specialist channels keep the meter honest.
Read and shareThe log changed the system
Observability is part of the experiment. Wall time and end-to-end outcomes beat a reassuring dashboard when the measurement path changes the target.
Read and shareAbout
I’m Alex Lennon (@embedded_iot). This notebook is engineering practice and insights in public — clocks, playbook, patterns, cases and starters. Steal what helps; measure your own results. AI-assisted; use at your own risk.
Public baseline v1.2.0 (2026-09-12) — token-disciplined, Codex-first, Cursor-compatible — baseline · v1.1 history. Updates land in the end-of-week review when the week earned them.
Content CC BY-SA 4.0 · code MIT · LICENSE