M1 Core Vocabulary

The seven terms and four intuitions of Stage 1. Source: the Agentic Operations Knowledge Map, Cluster A.

The M1 gate: you can explain, in your own words, why an agent's behavior is entirely a function of what's in its context window — and you can ship a reliable single-tool agent.

The seven terms

LLM: A next-token predictor trained on massive text. It doesn't "know" facts — it produces the statistically most plausible continuation. Why it matters: every capability and every failure traces back to this one mechanism.
Token: The unit an LLM reads and writes — roughly ¾ of a word, or a code fragment. Models think in tokens, not characters or concepts. Why it matters: cost, latency, and context limits are all measured in tokens.
Context: Everything the model can "see" right now: system prompt, instructions, conversation, retrieved data, tool outputs. Why it matters: the model has no memory beyond context — if it's not in context, it doesn't exist to the model.
Context window: The fixed maximum number of tokens the model can hold at once. Exceed it and earlier content is truncated or dropped. Why it matters: it's the hard physical boundary every agent design works around.
Context management = performance ⚠️: How you select, order, and compress what goes into the window directly determines answer quality. A model with great context beats a "smarter" model with cluttered context. Why it matters: the single highest-leverage skill in the field, and the most underrated by beginners. See Anthropic on context engineering.
Agent: An LLM placed in a loop with tools and a goal, deciding its own next action rather than answering once. Why it matters: the loop + tools is what turns a chatbot into something that can do work. See Anthropic's Building Effective Agents.
Tool call (function calling): The model emits a structured request (name + arguments); your code runs it and feeds the result back into context. Why it matters: the only way an LLM reaches beyond its training data into the real world. Mechanics: tool use docs.

The four critical intuitions

The model is stateless. Every turn, you rebuild its entire world from scratch in the context window.
"Garbage in, garbage out" is literal. Vague context produces confidently wrong output.
⚠️ More context is not better context. Irrelevant tokens actively degrade reasoning — measured as context rot.
A tool call is just structured text the model proposes. Your code is what makes it real — and safe.

The three M1 exercises

Build a single-tool agent and log the full context sent on every turn — read what the model actually "sees."
Deliberately overflow a context window with junk, then measure how answer quality degrades.
Rewrite one vague prompt three ways and compare outputs to feel how phrasing moves results.