Architecture · Economics · Work
Many Cheap Agents, or One Smart One?
I ran the bakeoff before writing a word. The question was the wrong one.
Lab report · all 72 runs, rawThe system around the algorithm
AI value isn’t the model. It’s the system around it.
I explore the people, work, architecture, governance, and economics that turn AI capability into better business and better work — then I build it myself, and show my work.
The system around the algorithm
The model is one part of the system.
Durable value depends on what surrounds it.
The system around the algorithm
Five dimensions, one model.
Hover or focus a dimension to read the question it asks. Select one to jump to field notes tagged with it.
The model in the center is deliberately the smallest, quietest node here — that’s the whole argument.Field notes
What I’m thinking about.
People · Work · Architecture · Governance · Economics
The System Around the Algorithm
Why good models still fail in real organizations.
Governance
Guardrails Are Not Governance
One controls behavior. The other defines accountability.
Architecture · Governance
Hallucination Is a Feature, Not a Bug
Not an argument for tolerating it. An argument for building the system that catches it.
Economics
The ROI Number in Your AI Business Case Is Probably Wrong
A framework for what an enterprise AI system is actually worth.
Governance · Architecture
If Your AI Can’t Show Its Work, Don’t Trust the Number
Why provenance belongs in the architecture, not bolted on after.
Architecture · Economics
Stop Using Your Biggest Model for Everything
My compute tiering rule.
Architecture · Governance
I Make My AI Fight Itself Before I Trust It
Builder, critic, and contrarian agents, explained.
Architecture · Context
Prompt Engineering Isn’t Dead. It’s Just Not Enough.
Context engineering is the bigger job.
Nothing tagged with this dimension yet.
What’s running
I write about systems I actually run.
Experiments, agents, evaluation loops, memory systems, and working demos. Some are stable. Some are still being broken on purpose.
What the colors and lines mean
Every dot and line below is real — nothing here is decorative.
- red — load-bearing, composed by several
- grey — composes or is composed
- ink — standalone, not wired in yet
- line — a real composes relationship
- outer ring — click a demo to open it
Playbooks
Reusable thinking, not one-off posts.
-
The Gauntlet Loop
Generate, challenge, revise, re-test, ship. The evaluation method behind almost everything else here.
-
The AI Agent Lifecycle
Context, plan, execute, evaluate, challenge, learn, deploy, monitor — the stages I actually build against.
-
Compute Tiering
Routing work between deterministic code, small models, and reasoning models by complexity, risk, and cost — not defaulting to the biggest one.
-
Falsifiable AI
Writing down what would prove a system wrong before running it, and building ground truth you can’t cheat.
-
Shared Ontology
One vocabulary every agent uses, so “the same data” actually means the same thing.
About
Twenty years in enterprise technology, one recurring question.
I work at the intersection of AI strategy, enterprise architecture, and transformation. A lot of my work is figuring out the system around the AI: the people, workflow, architecture, governance, and economics that determine whether it actually creates value.
LinkedInBetter business. Better work.
That’s what the system around the algorithm is for.
© 2026 Gordon Chan