Where a model stops answering and starts doing. Titles are on the spines; the notes underneath are the three tellings, in short — for a ten-year-old, for the architect, and for the boardroom.
6 concepts on this shelf · Beginner → Advanced · about 2 min read
Kitchen table — Not just answering a question — it can go and do the steps, then check whether it worked.
Whiteboard — A model plus a loop plus tools plus state. The model is the smallest part; the harness around it is where reliability is won or lost.
Boardroom — The shift from 'AI that advises' to 'AI that acts'. That is where the value is, and also where the governance questions begin.
AI agents, agentic AI, autonomous agents, agent harness, harness engineering, tool use
Kitchen table — Should it decide each step as it goes, or follow a map you drew for it? Usually a bit of both.
Whiteboard — A loop decides at runtime; a graph decides at design time. Loops wander, graphs go brittle. Real systems put a graph around the parts that must not vary.
Boardroom — How much freedom the system gets. More freedom, more capability and more variance — that dial is an executive decision, not a technical one.
agent loop, agent graph, orchestration, state machine, workflow, control flow
Kitchen table — Giving it hands. Now it can look things up, do the sum, or send the form — instead of just talking about it.
Whiteboard — Function calling with schemas, plus MCP as the emerging standard for exposing them. Naming, error messages and idempotency decide whether the model uses a tool correctly.
Boardroom — This is the integration layer — how AI reaches your actual systems. It is also exactly where access control has to be enforced.
tool use, function calling, MCP, Model Context Protocol, API integration, tool design
Kitchen table — What it still remembers about you tomorrow, and what it is better off forgetting.
Whiteboard — Working context, episodic notes and durable facts are three different problems. Writing is easy; deciding what deserves to persist, and expire, is the hard half.
Boardroom — Memory makes the experience feel personal — and turns every stored fact into a privacy and retention question you now own.
agent memory, long-term memory, context, personalisation, retention, state
Kitchen table — Like training a dog. Try something, get a treat or nothing, and slowly work out what earns treats.
Whiteboard — Agent, environment, reward, policy. Credit assignment over long horizons is brutal, reward hacking is real, and simulation quality decides everything.
Boardroom — Right when there is a clear score and cheap practice — pricing, routing, scheduling. Wrong when a mistake in the real world is expensive.
reinforcement learning, RL, reward function, policy, exploration, bandits, Q-learning
Kitchen table — People look at two answers and pick the better one. Do that a lot and the machine learns what people prefer.
Whiteboard — Reward models over pairwise preferences, PPO and the DPO family, plus constitutional and rubric-based variants. Alignment is a data problem wearing a maths hat.
Boardroom — How a raw model is turned into something safe and useful to ship. Whose preferences went in is a question worth asking your vendor.
RLHF, DPO, preference tuning, reward model, alignment, human feedback