Vocabulary

The working vocabulary of the field — enough to read a blog post, a job listing, or a framework’s docs without stopping. Terms marked ★ are the ones that come up constantly; internalize those first. Terms grouped by theme, because the grouping is itself a map of the field. This page is also a reference: come back when a later section uses a word you’ve lost.

The loop

• Agent ★ — a language model called in a loop, choosing actions and reacting to their results. An agent, unlike a workflow, decides its own steps.
• Harness ★ — the ordinary software around the model: it assembles input, executes the model’s chosen actions, and decides when the loop stops. When people say “the agent did X,” half the time the harness did X.
• Turn — one pass through the loop: one model call plus whatever actions follow from it.
• Trajectory ★ — the full record of an agent’s run: every model output, tool call, and result from start to finish. What you read when you debug.
• System prompt ★ — the standing instructions the harness puts at the top of every model call: role, rules, and constraints that apply to the whole task.
• Token ★ — the unit models read and write, roughly three-quarters of an English word. Context limits, pricing, and latency are all denominated in tokens.
• Tool call ★ — the model’s way of acting: it emits a structured request (“call search with query X”), the harness executes it and returns the result.
• Inference — running the model to get output. “Inference cost” is the pay-per-token bill for every call.

Context

• Context window ★ — the maximum amount the model can read in a single call. Everything the model knows about your task must fit in it, every turn.
• Context engineering ★ — the discipline of choosing and arranging what goes into the window: instructions, history, retrieved material. Successor to “prompt engineering,” widened from wording one message to curating everything the model sees.
• Compaction — shrinking a long transcript to fit the budget by summarizing or dropping old turns. Compaction, unlike truncation, tries to preserve meaning rather than just cutting from the top.
• Retrieval / RAG — fetching only the relevant slice of a large corpus into context at the moment it’s needed, instead of pre-loading everything. (RAG: retrieval-augmented generation.)
• Context rot — the measured decline in a model’s attention quality as its window fills; the reason a bigger budget doesn’t mean spend freely.
• Lost in the middle — the specific failure where models recall the start and end of a long context but miss what’s buried in the middle.
• Memory — durable knowledge across sessions, engineered by having the agent write notes it can re-read later. Memory, unlike context, survives the end of the conversation.

Tools

• Tool schema — a tool’s machine-readable interface: name, description, parameters. The only part of your tool the model ever sees.
• MCP ★ — Model Context Protocol, the open standard for packaging tools so any agent can connect to them; USB for agent tools.
• Sandbox — an isolated environment where an agent’s actions can’t damage anything real. Where you run code an agent wrote.
• Least privilege — granting an agent only the capabilities its task needs — read access, not write; one directory, not the disk. The first line of defense against both bugs and hijacking.
• Structured output — constraining the model to emit a machine-parseable format (usually JSON) so the harness can act on it without guessing.

Evaluation

• Eval ★ — a set of real test tasks with a grading method, run repeatedly to measure whether an agent works. An eval, unlike a unit test, yields a score with variance, not pass/fail.
• Golden set — the curated tasks-with-known-good-answers an eval grades against.
• LLM-as-judge — using a model to grade another model’s output against a rubric; the standard way to score outputs too open-ended for exact matching.
• Pass@k — the odds of at least one success in k attempts; the standard way to report success on tasks where outputs vary run to run.
• Trace — the logged trajectory of a production run, with timing and cost attached; the raw material of debugging and production evals.
• Regression — a change that improves the case you were staring at and quietly breaks others; the failure evals exist to catch.

Scale and cost

• Prompt caching ★ — reusing the computation for a context prefix the model has already read, making repeated turns dramatically cheaper and faster. The single biggest cost lever in agent engineering.
• Model routing — sending each step to the cheapest model that can handle it: a small model for extraction, a large one for reasoning.

Coordination

• Orchestrator — the agent (or plain code) that decomposes a task and delegates pieces to other agents.
• Subagent — an agent spawned to do one scoped piece of work in its own fresh context window, returning a distilled result.
• Handoff — the packet of context one agent passes another. Handoffs are where multi-agent systems lose information, the way phone-tag loses nuance.

Failure modes

• Hallucination — the model confidently asserting something false. A model quality problem — contrast with prompt injection, which is an attack.
• Prompt injection ★ — malicious instructions hidden in content the agent reads (a web page, an email) that hijack its behavior. The signature security problem of the field, because models read instructions and data through the same channel.
• Doom loop — an agent repeating a failing action with minor variations, burning tokens without progress; the agentic equivalent of an infinite loop.