Glossary
65 terms from "Delegating to an AI agent: a practical introduction." Look them up when you get stuck; the first mention in the text carries a hover definition.
| Term | Definition | Source |
|---|---|---|
| agent task | A request with a produced deliverable, a bounded scope, and a way to check completion, which is what makes it worth delegating rather than just asking. | Anthropic: Building Effective AI Agents |
| chat question | A request whose output is an answer you read and judge by understanding, with no fixed deliverable and no criteria to check it against. | OpenAI Academy: AI Foundations |
| delegable task | A request you can hand off because someone other than the agent can tell whether it succeeded. | OpenAI Academy: Agents and Workflows |
| temp worker | The course's image for an agent: a helper who can leave the desk, look things up, draft, and return, and who will guess when the instructions are vague. | Anthropic: Building Effective AI Agents |
| chatbot | An AI that answers turn by turn without taking actions for you, so a mistake stays a wrong answer rather than becoming a wrong action. | Anthropic: Building Effective AI Agents |
| clear goal | One sentence in the ask that states what finished looks like. | Anthropic: Building Effective AI Agents |
| bounded scope | A stated limit on what the agent should not do, given at the same time as the goal. | NimbleBrain: AI Agent Failure Modes |
| verification method | The checklist you plan to compare the result against, decided before the agent starts. | OpenAI Academy: Agents and Workflows |
| verifiable result | An output whose correctness a human can settle by comparison, without asking the agent again. | OpenAI Academy: Agents and Workflows |
| bounded output format | A stated shape for the deliverable — rows, columns, length, allowed values — that lets you check completeness by counting. | OpenAI API: Prompt Engineering |
| finish line | The point stated in the ask at which the job is over, which a chat-style request never supplies. | OpenAI Academy: AI Foundations |
| single deliverable | The one artifact an agent task returns, so that checking has a definite object. | OpenAI Academy: Agents and Workflows |
| clear ask | A structured request carrying task, requirements, boundaries, and acceptance criteria. | OpenAI API: Prompt Engineering |
| four-part ask | The template of task, then requirements, then boundaries, then acceptance criteria, written in that order. | Claude Platform: Prompting Best Practices |
| one action | The single operation named in the task line, with no hidden steps folded in. | OpenAI API: Prompt Engineering |
| requirements | The properties the output must have: format, length, tone, sources, structure. | Claude Platform: Prompting Best Practices |
| boundaries | The part of the ask stating what the agent must not do, which heads off unwanted side trips. | NimbleBrain: AI Agent Failure Modes |
| four pieces of information | The lesson's framing of what an agent actually needs — as opposed to more words. | OpenAI API: Prompt Engineering |
| vague wish | A request like "make my résumé better" that names a direction but no action, shape, limit, or standard. | OpenAI API: Prompt Engineering |
| do not read minds | The lesson's reminder that an agent has only your prompt, not your intent. | OpenAI API: Prompt Engineering |
| most common pattern | What an agent falls back on when the prompt leaves a gap: the interpretation seen most often in training, not your case. | OpenAI API: Prompt Engineering |
| fence off unwanted behavior | What the boundaries step does: closing off actions before the agent has a chance to choose them. | NimbleBrain: AI Agent Failure Modes |
| measurable outcome | A property of a rewritten output that can be confirmed by looking, such as a number appearing in a bullet. | Claude Platform: Prompting Best Practices |
| optimize for the wrong goal | What happens when the checklist appears before the task, so the agent aims at the criteria instead of the job. | Claude Platform: Prompting Best Practices |
| verification | Checking each acceptance criterion against evidence, as opposed to reading the output. | OpenAI Academy: Agents and Workflows |
| verification table | A three-column record of criterion, evidence, and pass or fail, used to check a deliverable. | OpenAI Academy: Agents and Workflows |
| verification checklist | The short list of criteria, each specific enough to be checked without guessing, that you take into a review. | OpenAI Academy: Agents and Workflows |
| acceptance criteria | The conditions the output must meet, written before the agent starts and reused as the test afterwards. | OpenAI Academy: Agents and Workflows |
| acceptance criterion | A single condition from the checklist, checked on its own against its own evidence. | OpenAI Academy: Agents and Workflows |
| criterion, evidence, and pass or fail | The three columns of the verification table, naming the standard, where you looked, and the verdict. | OpenAI Academy: Agents and Workflows |
| Reading is passive | The course's contrast between reading an output and verifying it: reading takes what is offered, verification asks a question of it. | OpenAI Academy: Agents and Workflows |
| confident tone | The assured delivery an agent uses regardless of whether the content is right, which is why tone carries no information about correctness. | Anthropic: Building Effective AI Agents |
| looks about right | The second trap in checking: moving on because the shape of the output matches what you expected. | OpenAI Academy: Agents and Workflows |
| not a reliable judge | The reason asking the agent whether it followed your instructions does not count as a check. | Claude Platform: Prompting Best Practices |
| calendar date | The countable form a deadline criterion demands, as opposed to a relative phrase like "ASAP" or "next sprint". | OpenAI Academy: Agents and Workflows |
| specific owner | The criterion that an action item names a person, not a team. | OpenAI Academy: Agents and Workflows |
| countable property | A criterion restated so it can be settled by inspection, replacing a subjective one. | OpenAI Academy: Agents and Workflows |
| hallucination | An invented fact, name, date, citation, or action produced with the same confidence as correct output. | AI/TLDR: Common AI Agent Failure Modes |
| stale fact | Information that was true once but is no longer current, because the agent's knowledge has a cutoff date. | AIToolsUsageGuide: Why AI Gives Wrong Answers |
| scope shrink | When an agent quietly does less than you asked, narrows the request, or swaps in an easier task without telling you. | AI/TLDR: Common AI Agent Failure Modes |
| non-answer | A response that looks full — background, definitions, general advice — but never addresses the question asked. | AI/TLDR: Common AI Agent Failure Modes |
| scope creep | When an agent expands past its mandate and adds work you did not request — the opposite direction from scope shrink. | NimbleBrain: AI Agent Failure Modes |
| four mistake types | The course's checking framework: hallucination, stale facts, scope shrink, and non-answer, each with a different tell. | AI/TLDR: Common AI Agent Failure Modes |
| plausible-looking continuation | What the model actually produces at every step, which is why a hallucination is not a special malfunction. | AI/TLDR: Common AI Agent Failure Modes |
| made-up citation | A source, section, or quotation the agent produces that does not exist in the material. | AI/TLDR: Common AI Agent Failure Modes |
| invented detail | A fact the agent supplies that you never gave and no source contains, such as a time added to a party post. | AI/TLDR: Common AI Agent Failure Modes |
| cutoff date | The point past which the agent's training knowledge stops, which is what makes prices, policies, and versions go stale. | AIToolsUsageGuide: Why AI Gives Wrong Answers |
| live source | A current reference you require in the ask so a time-sensitive claim can be checked against the world rather than memory. | AIToolsUsageGuide: Why AI Gives Wrong Answers |
| compare the claim against the source | The quick check for hallucination: locating the claim in the material rather than judging it by how it reads. | AI/TLDR: Common AI Agent Failure Modes |
| quietly narrows the task | How scope shrink presents: the agent answers one of three questions, summarizes instead of categorizing, or drops a boundary, without flagging it. | NimbleBrain: AI Agent Failure Modes |
| easier neighboring question | The question an agent answers instead of yours when yours is hard, producing a non-answer. | AI/TLDR: Common AI Agent Failure Modes |
| different fingerprint | The reason the four types are worth separating: each leaves its own trace and needs its own check. | AIToolsUsageGuide: Why AI Gives Wrong Answers |
| feedback loop | The designed-in cycle of reviewing output, saying what to fix, and checking the next draft against the same criteria. | Anthropic: Building Effective AI Agents |
| correction loop | The four-step cycle: compare the output to the criteria, describe the failure specifically, let the agent revise, verify the new output against the same criteria. | Anthropic: Building Effective AI Agents |
| mistake log | A record of the original ask, the mistake type, how you spotted it, and the corrected result. | Anthropic: Building Effective AI Agents |
| specific correction | A correction that names the location, the error, and the standard, so the agent can revise without guessing. | Anthropic: Building Effective AI Agents |
| vague correction | A correction like "fix this" that leaves the agent room to guess again, lengthening the loop. | Anthropic: Building Effective AI Agents |
| please do better | The lesson's example of a correction that carries no information about what to change. | Anthropic: Building Effective AI Agents |
| the location, the error, and the standard | The three things a correction must carry: where it is wrong, what is wrong, and what right looks like. | Anthropic: Building Effective AI Agents |
| spot, report, revise, verify | The four moves of the loop that make delegation reliable. | Anthropic: Building Effective AI Agents |
| human feedback | The input you supply at checkpoints, which is what keeps an autonomous agent from drifting off course. | Anthropic: Building Effective AI Agents |
| checkpoints | The points where you stop the work and review it, rather than waiting for the agent to finish everything. | Anthropic: Building Effective AI Agents |
| ground truth | Verifiable information from the environment that an agent checks against at each step instead of assuming its plan worked. | Anthropic: Building Effective AI Agents |
| revised output | The draft that comes back after a correction, which goes through the same criteria rather than straight to use. | Anthropic: Building Effective AI Agents |
| first mistake-log entry | The course's closing exercise: delegate a small real task, check it, and record the mistake type, how you spotted it, and the corrected result. | Anthropic: Building Effective AI Agents |