Agentic AI
AgentsA design stance in which AI systems pursue goals over multiple steps rather than producing one output per request.
- agentic systems
- agentic workflows
Loading the homepage
The vocabulary of AI engineering, defined the way an engineer would want it: what the thing actually is, the mechanism underneath it, when it matters, and the misconception that trips people up most often.
Search the list, filter by section, or jump straight to a letter. Every entry links out to the related terms it depends on and to the articles that put it to work.
34 terms
A design stance in which AI systems pursue goals over multiple steps rather than producing one output per request.
A system that uses a language model to choose actions, executes them through tools, and repeats until a goal is met.
The operation that lets each token in a sequence weigh every other token, producing context-dependent representations.
Having a model produce intermediate reasoning steps before its answer, which improves accuracy on multi-step problems.
The maximum number of tokens a model can attend to in one request, covering the prompt and the generated response together.
Training a smaller model to reproduce a larger one’s behaviour, trading some capability for much lower cost and latency.
Numeric vectors that place text in a space where distance approximates similarity of meaning.
Measuring an AI system’s output against a fixed dataset and scoring criteria so changes can be compared rather than guessed at.
Supplying a handful of worked examples in the prompt so the model infers the task pattern without any training.
Continuing training on a curated dataset so a model adopts a behaviour, format or domain style by default.
Declaring functions with typed parameters so a model can respond with a structured request to invoke one.
Deterministic checks around a model that constrain what may enter it and what may leave it.
Output that is fluent, confident and wrong — invented facts, citations or APIs presented in the same tone as correct ones.
A deliberate pause where a person reviews, approves or corrects an AI system’s proposed action before it takes effect.
Running a trained model to produce output — the phase where latency, throughput and cost are actually incurred.
A neural network trained on large text corpora to predict the next token, which in practice yields broad language and reasoning ability.
An open protocol that standardises how AI applications connect to external tools, data sources and prompts.
An architecture where several specialised agents, each with its own context and tools, cooperate on one task.
The layer that coordinates steps, state, retries and hand-offs across the model calls and tools that make up an AI workflow.
Designing the instructions, examples and context supplied to a model so its output is reliable enough to build on.
An attack in which text the model reads as data is treated as instruction, redirecting the system’s behaviour.
Storing model weights at lower numeric precision so a model needs less memory and less bandwidth to run.
An agent loop that interleaves a reasoning step with an action, so each tool result informs the next decision.
A second scoring pass that reorders retrieved candidates by judging each one jointly with the query.
Fetching relevant documents at request time and placing them in the prompt so the model answers from your data rather than its training.
Retrieval that matches on meaning by comparing embeddings, rather than on the presence of shared keywords.
Delivering a model’s output token by token as it is generated, so the interface responds immediately instead of waiting.
Forcing a model’s response to conform to a schema, so downstream code can parse it instead of interpreting prose.
The operator-supplied instruction block that establishes a model’s role, constraints and output conventions for a conversation.
A sampling parameter that flattens or sharpens the model’s next-token probability distribution, trading consistency against variety.
The unit a model actually reads and writes — a sub-word fragment, not a character or a whole word.
The loop in which a model requests an external capability, the application runs it, and the result re-enters the context.
The neural network architecture behind modern language models, built on self-attention rather than recurrence.
A store that indexes embeddings so nearest-neighbour queries stay fast as the collection grows.
The products behind the vocabulary — vector stores, agent frameworks, evaluation harnesses and the rest.
When two terms are presented as rivals — RAG against fine-tuning, for instance — this is where the trade-off is worked through.
Definitions get you the vocabulary. These builds show the same ideas running in real code.