The playbook is a map of where humans anchor the system.
24 modules, 5 groupings. The faint lines appear on hover — that’s where the thesis folds back on itself.
Origin · 3 modules
The Entropik Thesis
Three principles I ended up writing down because they were the shape of the realisations I kept having, platform after platform. I didn't start with them — I arrived at them, and this is the short version of why.
The Demon Principle
A thermodynamic analogue for why the human in the loop is load-bearing, not friction. Arrived at the long way — by watching three platforms get strange when I tried to remove the human.
The 6× Gap
I spent too long chasing model upgrades before noticing the harness was the thing doing the work. The research caught up in 2026 — and looking back at my older platforms, the scars line up with the findings.
Structure · 6 modules
Event Sourcing
I tried CRUD first. Of course I did — it's what you reach for. This is what taught me why it doesn't work for AI-first platforms, and what I ended up using instead.
Skills over Controllers
I wrote controllers for years. The pattern that replaced them was less obvious than I'd like it to have been — and this is the version of that story I wish someone had told me earlier.
Triple Output on Every Feedback Interaction
I had all three outputs — updated state, training signal, audit trail — but I'd built them as three separate systems. Which, it turns out, is almost as bad as not having them at all.
The Six Pillars
The six architectural shifts that separate an AI-first platform from a domain system with AI features bolted on. Not invented — noticed, after enough repetitions that the pattern was no longer deniable.
The Three-Layer Stack
Infrastructure provides context. Skills provide capability. Orchestration provides experience. Keeping these three separate is what lets any one of them evolve without dragging the other two with it.
The Feedback Loop
Accept, modify, reject. One pattern, three signals, three simultaneous outputs. I kept collapsing this into "we have a like button" and paid for the collapse for years.
Harness · 5 modules
Harness Anatomy
What a harness actually contains, once you stop thinking of it as glue around a model and start thinking of it as the operating system the model runs on. The parts I now reach for deliberately, after the third or fourth time I rebuilt one from scratch.
Optimisation Loops
The difference between a tool and a platform is whether feedback closes a loop. I spent years building tools — I'm still learning what it takes to close the loop properly.
The Subtraction Principle
Mature harness engineering is a craft of subtraction as much as addition. I got good at adding structure. Getting good at removing it has been harder, and the difference is where the leverage is.
Harness Your Development Agent
The harness I thought mattered was the one around the production agent. The one that actually compounds is the one around the agent that builds the production agent. This is the realisation I'm living through right now, writing this module.
Execution Contracts
Five bindings that turn an agent call from a black box into composable infrastructure. Getting this shape right is what made the difference between "AI feature" and something I could actually operate.
Practice · 6 modules
Autonomous Patterns
Five patterns I now trust when I build systems that act without asking permission. Each one is a rule I broke first, regretted, and eventually worked out the shape of.
Multi-Tenancy and Verticals
Tenants and verticals are independent axes. I kept collapsing them into one until the configuration sprawl made it obvious why I shouldn't. This is the shape that held up.
Context Assembly
The discipline that replaces data modelling when you're building AI-first. For years I kept asking what should we store — the right question is what should the AI see.
Memory Architecture
Six memory types, not one. I kept writing platforms as if "the database" was the whole memory story, and being surprised when the platform felt amnesic. This is the taxonomy I should have started with.
Skills Architecture
What a skill file actually looks like once you stop treating markdown as notes and start treating it as the capability itself. The shape I've converged on after more rewrites than I'd like to admit.
Boundary Skills
The difference between a tool and a cognitive hub is whether intelligence flows in and out across the perimeter. I kept building tools and wondering why they felt walled in.
Ground · 4 modules
The Eval Harness
If you can't measure it, you can't optimise it. If you can't optimise it, you don't have a platform — you have a demo. I shipped demos for a while before this clicked.
The Platform Audit
A 12-section checklist I now run on every AI platform I look at, including my own. Designed to surface architectural gaps honestly, not feature gaps.
Observability
Traditional APM measures response time and error rates. Agents need traces, cost attribution, and quality proxies — and I had to learn that the hard way after the first month I ran AI in production without them.
Security and Trust
Traditional RBAC and encryption aren't enough. Agents are non-deterministic multi-step actors with tool access — and they need a different security posture. These are the five boundaries I now think in.