Every term you’ll meet in this guide, in one place, in plain language. Bookmark it.
Nothing here is new if you’ve read Part 0 — this just pins it all down so a stray bit
of jargon never derails you later.
┌───────────────────────── the HARNESS (everything around the model) ─────────────────────────┐
TASK │ ┌──────────┐ context ┌───────┐ tool request ┌───────────┐ result │ ANSWER
────► │ │ build │ ───────────► │ MODEL │ ───────────────► │ run │ ─────────┐ │ ─────►
│ │ context │ └───────┘ │ the tool │ │ │
│ └──────────┘ ▲ └───────────┘ │ │
│ │ └──────────────── feed result back ──────────┘ │
│ │ (this arrow is the LOOP) │
│ │ MEMORY │ history · state · context │
│ every step is written to an EVENT LOG and CHECKPOINTED so a crash can resume │
└───────────────────────────────────────────────────────────────────────────────────────────────┘
The model is the small box in the middle. Everything else in that outer frame is the
harness — and building that frame, piece by piece, is the whole guide.
Term
In plain language
First seen
Model / LLM A function that turns input text into output text. No memory, no hands, runs once per call.
0.1
Prompt The text you send the model on a given call.
0.1
Context All the text the model sees this call — the prompt plus whatever history the program chose to include.0.1
Token The unit models read and write in — roughly a word-piece. Context size and cost are measured in tokens.
2.3
Term
In plain language
First seen
Tool A plain function the program will run on the model’s behalf.
0.2
Tool call / request The structured message the model emits to ask for a tool — data, not prose.
0.2
Tool result What the tool returns, fed back to the model as more text.
0.2
Safe vs. dangerous tool Safe tools read the world; dangerous tools change it and need guarding.
0.2
Term
In plain language
First seen
Agent A loop around a model that keeps calling it until the task is done.
0.3
Loop The code that repeats call → check → maybe run a tool → repeat.
0.3
Step One pass through the loop — one model decision, maybe one tool run.
1.1
Turn A related exchange: the model’s request plus the tool result that answers it.
1.1
These arrive in Part 2. You don’t need them yet — they’re here so the words are never
strangers when they show up.
Term
In plain language
First seen
Harness Everything wrapped around the model: context-building, tools, safety, memory, durability. The star of this guide.
1.1
Event log An append-only record of everything that happened, in order.
2.1
Checkpoint A saved snapshot of progress so a crash can resume instead of restarting.
2.1
State The agent’s running summary of what matters so far.
2.3
Policy A gate that inspects every tool call before it runs.
2.2
Sandbox An isolated place to run untrusted, model-written code safely.
2.2
Routing / handoff Sending a task to the right specialist agent, with its context.
2.4
Supervisor An agent that fans work out to sub-agents and merges their results.
2.5
Suspend / resume Pausing a workflow (e.g. to await a human) and continuing later — even days later.
2.6
The model decides; the harness executes. Every term above is either part of the
model’s small job or part of the harness’s large one.