Glossary

Agent

A model wrapped in a loop, given tools and state, pursuing a goal it can check. Not defined by intelligence — by the ability to act and adapt. 01

Agent-Computer Interface (ACI)

How an agent perceives and operates a computer: screen reading, input injection, window and file management. 04

Agent loop

The repeating cycle: build context → decide → check permission → execute tool → observe → update state → stop or repeat. 03

Autonomy level

A six-step scale from tool (L0) to self-improving agent (L5), describing how much the system decides without a human. 01

Autoregressive

Generating output one token at a time, each new token conditioned on everything before it. 02

Backoff

Waiting longer between retries — usually with random jitter — so a struggling service isn't hammered. 09

Benchmark

A standard test set for comparing models or systems. Useful context; never a substitute for your own eval set. 08

Budget cap

A hard limit on tokens, money or time per task, enforced by the harness rather than requested in a prompt. 09

Chain

A fixed sequence of model calls with no branching. Predictable and cheap; not an agent. 01

Checkpoint

A serialized snapshot of agent state after a step, so work can resume rather than restart. 09

Chunking

Splitting documents into retrievable pieces — by heading, section or function, with slight overlap. Retrieval quality starts here. 05

Circuit breaker

After repeated failures from a dependency, fail fast for a cool-down period instead of retrying into a wall. 09

Citation / grounding

An answer carrying the specific source passages it came from. The practical test of whether retrieval worked. 05

Compaction

Compressing conversation history — recent turns verbatim, older ones summarized or extracted — to stay inside the window. 05

Context engineering

Deciding exactly what text occupies the context window on each call, and what is deliberately left out. 05

Context rot

Gradual drift in behavior as an agent works from summaries of summaries and the original constraint fades. 08

Context window

The fixed token budget for one model call: instructions, history, tools, retrieved text, and answer room. 02

Cost per completed task

Total tokens, retries and overhead divided by tasks actually finished successfully. The only cost metric worth a dashboard. 09

Dead-letter queue

A holding area for tasks that failed permanently, kept with their trace for human review instead of being dropped. 09

Delegation

The parent keeps ownership of the goal and spawns a bounded worker; the worker returns a result and ends. 07

Determinism

Same input, same output. Useful for testing; deliberately traded away when sampling for variety. 02

Embedding

A numeric vector representing meaning, so similar text lands near other similar text in vector space. 05

Episodic memory

A stored record of what happened — past sessions, attempts and outcomes — retrieved later by similarity or time. 05

Eval / golden set

A fixed set of real cases with programmatic end-state checks, run before every change to catch regressions. 08

Fine-tuning

Further training of a model on your examples to change its behavior. Rarely the fix for a context or tool problem. 10

Function calling

The model emitting a structured request — tool name plus JSON arguments — that the harness validates and executes. 04

Guardrail

A check on model input or output: schema validation, content filter, allowlist, or a rule that blocks an action outright. 04

Hallucination

Fluent, confident output that isn't true. A property of prediction, not a bug that a prompt can reliably remove. 02

Handoff

Ownership of the conversation moves to another agent or a human; the original agent stops driving. 07

Harness

Everything around the model that makes it an agent: loop, tools, permissions, state, context builder, tracing. 03

Human-in-the-loop (HITL)

Deliberate human decision points — approving a plan, an action, or an outcome — placed where mistakes are expensive. 06

Idempotency

Running an action twice has the same effect as once, usually via a unique key. The cure for double charges and duplicate emails. 09

Inference

Running a trained model to produce output. What you pay for per token. 02

Jailbreak

Coaxing a model past its own policies using role-play, encoding or emotional pressure. Distinct from prompt injection. 08

JSON schema

A machine-readable description of allowed input: types, required fields, enums, defaults. The tool's contract. 04

Latency (TTFT / end-to-end)

Time to first token, and total time until the user has the result. Agents multiply the second by their step count. 09

Least privilege

Giving each task the narrowest credential and reach it needs, for as short a time as possible. 09

LLM-as-judge

Using a model to score outputs against a rubric. Cheap and useful, but biased toward verbosity and sensitive to phrasing — calibrate on humans. 08

MCP (Model Context Protocol)

A standard way to expose tools, resources and prompt templates to any compatible client over JSON-RPC. 04

Memory (working / semantic / procedural)

The current context; distilled facts and preferences; and how-to knowledge stored as prompts, examples or code. 05

Mixture of experts (MoE)

An architecture where each token is routed to a few specialised sub-networks, giving large capacity at lower compute per token.

Model routing

Sending each request to the cheapest model that can handle it, escalating on low confidence or validation failure. 06

Orchestrator

The component that decomposes a goal, assigns work, monitors progress and merges results — code, not a model. 07

Permission gate

A checkpoint where the harness allows, asks about, or denies an action based on its risk tier. 04

Plan-and-execute

Generate a reviewable plan first, then work through it, replanning only on failure. 06

Prompt injection

Instructions hidden in content the agent reads — a web page, email, PDF or tool result — that the model may follow as if they came from you. Defeated structurally, not by prompting. 04

Prompt caching

Reusing the computed state of an unchanged prompt prefix to cut cost and latency on repeated calls. 09

Quantization

Storing model weights at lower precision (8-bit, 4-bit) to shrink memory and speed inference, at some quality cost.

RAG (retrieval-augmented generation)

Retrieving relevant passages and putting them in the context so the answer can be grounded and cited. 05

Reasoning model

A model that spends extra tokens thinking before answering. Better on multi-step problems; slower and pricier. 02

ReAct

Interleaving reasoning and acting: think, call a tool, read the result, think again. 06

Reflection

Having the model critique and revise its own output. Helps with checkable errors; weak on confident factual mistakes. 06

Regression

A case that used to pass and now fails. The reason a golden set exists. 08

Reranking

Re-scoring a broad set of retrieved passages with a stronger model to keep only the best few. Often the best quality-per-dollar upgrade. 05

Sandbox

An isolated environment for running model-generated code, ideally with no credentials, no network and a hard timeout. 04

State

The structured record of progress: messages, task list, results, budget, retries. Lives in code, not in the model's head. 03

Sub-agent

A worker instance with its own context and tool permissions, used to isolate work or run pieces in parallel. 07

Swarm

Peer-to-peer agents with no central controller, coordinating through messages or the environment. Research-first; hard to audit. 07

Temperature

The sampling knob: low means predictable and repetitive, high means varied and more error-prone. 02

Test-time compute

Spending more tokens or search at answer time — reasoning traces, multiple attempts, verification — instead of training a bigger model.

Token

A chunk of text, roughly 3–4 characters of English. The unit of context, cost and latency. 02

Tool

A function the model may request: a name, a when-to-use description, a strict input schema, and a small useful output. 04

Trace

The recorded sequence of model calls, tool calls, decisions and timings for one task. Your debugging interface. 09

Vector store

A database of embeddings with fast nearest-neighbour search. The index behind retrieval. 05

Workflow vs agent

A workflow follows paths you wrote; an agent chooses paths at runtime. Prefer the workflow where it fits. 01

No terms match that search.