M2 Core Vocabulary

The four clusters of Stage 2. Source: the Agentic Operations Knowledge Map. Sister sheet to M1 Core Vocabulary.

The M2 gate: you can design a reliable multi-step agent with structured prompting, external memory, and a chosen reasoning pattern — and explain why you made each choice, not just copy it.

Cluster A — Prompt engineering

Structured prompting: Clear role, constraints, output format, and few-shot examples — instead of a bare ask. Why it matters: the prompt is the part of the context you fully control; structure is how you spend that control. Technique catalog: prompt engineering docs.
Instruction hierarchy: System > developer > user — instructions carry different authority depending on where they sit, and ordering matters. Why it matters: it's the difference between rules of engagement and requests — and the thing injection attacks try to subvert.
Prompt injection ⚠️: Untrusted data placed in context can hijack instructions ("ignore previous instructions and…"). The model can't reliably tell data from directive. Why it matters: every retrieved document, tool output, and pasted page is an attack surface — and the problem is still unsolved. See Willison's series.

Cluster B — Memory types

In-context memory: What lives in the current window. Fast, but volatile and size-capped.
External memory: Stored outside the model — files, databases, vector stores — and pulled in on demand.
Episodic memory: Records of past runs and interactions the agent can recall to inform future behavior.
The memory intuition ⚠️: Memory is a retrieval-and-injection problem, not a "the model remembers" problem — you are always choosing what to reload. Why it matters: "add memory" is never a feature toggle; it's a design decision about what enters the window, when.

Cluster C — Reasoning & planning patterns

Chain-of-thought (CoT): Prompt the model to reason step-by-step before answering; linear. Origin: Wei et al., 2022.
ReAct: Interleave reasoning traces with actions — plan, act, observe, adjust, in a loop. The workhorse pattern for tool-using agents. Origin: Yao et al., 2022.
Tree-of-Thought (ToT): Explore multiple reasoning branches with self-evaluation, lookahead, and backtracking — for problems with many candidate paths. Origin: Yao et al., NeurIPS 2023.
The pattern intuition: Match the pattern to the problem: CoT for linear reasoning, ReAct for tool-driven tasks, ToT for search/exploration where the cost is justified.

Cluster D — Tool use & retrieval

Tool schemas: Clear names, typed arguments, tight descriptions the model can act on. Why it matters: the model chooses tools by reading their descriptions — tool descriptions are prompts.
Retrieval-Augmented Generation (RAG): Fetch relevant external chunks and inject them into context, so answers are grounded in real data rather than the model's parametric guesses.
RAG failure modes ⚠️: Bad chunking and weak retrieval poison answers more often than the model itself does.

The three critical intuitions

A reasoning pattern is a cost/quality tradeoff, not a magic upgrade. ToT can be many times more expensive than CoT.
Tool descriptions are prompts. The model picks tools based on how you describe them.
⚠️ Retrieval quality caps answer quality. You can't reason your way out of missing context.

The three M2 exercises

Build a ReAct agent that searches, reads, and answers a multi-hop question, logging each thought/action/observation.
Build a small RAG pipeline and deliberately break the chunking to see grounding collapse.
Add episodic memory so an agent recalls a prior session's outcome and changes its plan.