Topic Index

A thematic view of the dictionary. For the alphabetical list, see entries/. To return to the front page, see home.

Foundations

These are the load-bearing concepts. Most other terms reference one or more of these.

Agent — what an agent is, how it differs from a chatbot, and why Gibson’s Agency reads like a design document.
Embedding — meaning as a list of numbers.
Tool — the function call that lets an agent act in the world.
Naming — why the choice of names is structural, not cosmetic.

The architectural pieces of a running agent.

Gateway — the always-on coordinator process.
Sub-agent — delegated AI sessions for parallel or focused work.
Heartbeat — periodic, automated nudges that make agents proactive.
SOUL.md — the agent persona file as architectural pattern.
Aunties — specialized single-verb oversight agents that prevent any one component from accumulating unchecked authority.
The Lowbeer Question — who holds the authority to terminate, who executes it, and what happens when the principal is not available.

How an agent uses information beyond its training data.

RAG (Retrieval-Augmented Generation) — the dominant pattern for “AI that knows my stuff.”
Vector database — the storage and retrieval infrastructure for embeddings.
Fine-tuning — when to retrain a model, and when not to.

The shared protocols and runtimes that make agentic systems composable.

MCP (Model Context Protocol) — the open standard for connecting agents to tools.
Ollama — local LLM runtime for sovereignty and cost containment.

How models are built, sized, distributed, and accessed — and the strategic camps the industry has organized itself into.

Parameters — what the “26B” or “70B” in a model name actually means, and why it matters less than vendors imply.
Open source — AI models whose weights are published publicly. The hedge against vendor lock-in.
Closed source — AI models delivered only through vendor-hosted APIs. Where the frontier currently lives.

The ongoing reality of running an agent in production — the costs, the meters, and the feelings practitioners develop about them.

Token burn — the cost-rate concept. What is it costing?
Token anxiety — the capacity-bounded operational concept. Will it fit?
Token angst — the existential concept. Was it worth it?
Dusty Laptop — the minimum-viable hardware entry point. The old machine in the closet that suddenly has a use.

The practitioner’s craft of using AI assistants without breaking your own learning, your own thinking, or your own labor-market position.

English major — why the new bottleneck is specification, not syntax, and who that suddenly favors.
Can’t help you understand — the hard limit on what an AI can do for your comprehension.
Descartes was wrong — a philosophical reframing of what is happening when an AI agent (or a human) thinks.

The dictionary is a work in progress. Expected near-term additions:

Skill — packaged capabilities aimed at the agent.
Context window — the boundary of what a model can see at once.
Prompt / system prompt — the input layer of an agentic system.
Token — the unit of cost, throughput, and capacity.
Quantization — why a 70-billion-parameter model can fit in 42 gigabytes.
Hallucination — what it is, what it isn’t, and why “hallucination” is itself an imperfect name.
Local-first / sovereignty — running AI without sending data to a third party.
Model tiering — using different models for different tasks to control cost.
Approval gating — how to require human consent for sensitive agent actions.
Provenance — knowing where an agent’s output came from.

If a term you wish were here is missing, open an issue and the maintainer will add it.

Maintained by Matthew D. Langenkamp / 雷邁德.