3.2 Glossary

Plain-language definitions for every term the guide uses, with a link back to where each one is introduced. Skim it now to lock things in, or bookmark it and return whenever a word trips you up.

The whole system, in one picture

        ┌───────────────────────── the HARNESS (everything around the model) ─────────────────────────┐
        │                                                                                              │
  TASK  │   ┌──────────┐   context    ┌───────┐   tool request   ┌───────────┐   result                │  ANSWER
  ────► │   │  build   │ ───────────► │ MODEL │ ───────────────► │  POLICY   │ ─► run / sandbox / block │ ─────►
        │   │ context  │              └───────┘                  │   gate    │          │              │
        │   └──────────┘                  ▲                      └───────────┘          │              │
        │        ▲                        └──────────────── feed result back ───────────┘              │
        │        │                                    (this arrow is the LOOP)                          │
        │   ┌────┴─────┐        every step → EVENT LOG + CHECKPOINT (durable; enables resume & suspend) │
        │   │  MEMORY  │        route to specialists · fan out to sub-agents · suspend for a human      │
        │   └──────────┘                                                                                │
        └───────────────────────────────────────────────────────────────────────────────────────────────┘

The model is the small box. Everything else is the harness.

A–Z

Agent — A loop around a model that keeps calling it until a task is done. Not a chatbot and not magic. (0.3)

Checkpoint — A saved snapshot of a workflow’s progress, written durably so a crash can resume instead of restarting. (2.1)

Context — All the text the model sees on a given call: the prompt plus whatever history the harness chose to include this step. Rebuilt every step, not a growing log. (0.1, 2.3)

Context engineering — The production discipline of deciding what goes into context — retrieval, summarization, trimming — to keep it small and relevant. (2.3)

Durable execution — Running a workflow so that every step is checkpointed and a crash resumes exactly where it left off, with no repeated side effects. (2.1)

Event log — An append-only record of everything that happened, in order, stored somewhere that survives a crash. The source of truth for history. (2.1)

Handoff — A typed transfer of a task from a router to a specialist agent: intent plus extracted details, not vague free text. (2.4)

Harness — Everything wrapped around the model: context-building, tools, policy, memory, durability, routing, supervision, approval. The subject of this whole guide. (1.1)

History — Everything that has happened in a task, in full. Kept in the event log; not fed to the model wholesale. (2.3)

Human-in-the-loop (HITL) — Requiring a person’s decision before an action proceeds, modeled as a durable suspend/resume rather than a blocking wait. (2.6)

Idempotency — The property that doing something twice has the same effect as doing it once. What stops a resumed workflow from re-sending an email. (2.1)

LLM (large language model) — A function from text to text. No memory, no hands, runs once per call. The engine an agent is built around. (0.1)

Loop — The code that repeats call the model → check for a tool → maybe run it → repeat, until the model answers without asking for a tool. (0.3)

Policy — A gate that inspects every tool call before it runs and decides: allow, sandbox, or block. The harness’s safety boundary. (2.2)

Prompt — The text you send the model on a given call. (0.1)

Routing — A control-plane decision: read a task’s intent and pick the specialist agent that should handle it. (2.4)

Sandbox — An isolated runtime with strict limits (time, memory, no network/filesystem) where untrusted or model-written code can run without doing harm. (2.2)

Side effect — An action that changes the outside world — sends an email, charges a card, deletes a row. The reason crashes and retries are dangerous. (1.2)

State — A running summary of the older history: the agent’s compact working memory. (2.3)

Step — One pass through the loop: one model decision, possibly one tool run. (1.1)

Structured error — An error returned as data ({ error, reason }) rather than a crash, so the model can read it and self-correct. (2.2)

Sub-agent — An agent spawned by a supervisor to handle one piece of a larger task, in isolation and often in parallel. (2.5)

Supervisor — An agent that plans, fans work out to sub-agents in parallel, collects their results, and synthesizes an answer — surviving partial failure. (2.5)

Suspend / resume — Pausing a workflow into a durable “suspended” state (e.g. awaiting approval) and continuing it later, even after a restart or a days-long wait. (2.6)

Token — The unit models read and write in, roughly a word-piece. Context size and cost are measured in tokens. (2.3)

Tool — A plain function the program runs on the model’s behalf. The model can only request it. (0.2)

Tool call / request — The structured message the model emits to ask for a tool — data, not prose. (0.2)

Turn — A related exchange within a task: the model’s request plus the tool result that answers it. (1.1)

Workflow — The whole durable, multi-step process of completing a task. “Agent systems are workflow systems.” (2.1)