Glossary · 114 terms
AI & LLMs → Agents glossary
Every new word in the AI & LLMs → Agents course, in plain English, explained on a card the first time it comes up.
#
- .skip
- test marker that skips a test, so its failure no longer shows (it.skip is the same thing)Unit: Reward hacking & specification gaming
A
- admin rights
- full control over a system: the most powerful and risky kind of accessUnit: Checkpoint: Advanced
- alert
- an automatic message that tells people something needs attention nowUnit: Evals & guardrails
- allow-list
- a list of the only things permitted; anything not on it is blockedUnit: Evals & guardrails
- analytics
- analysing data to answer business questions; analysts are the people who do itUnit: AI from zero
- API
- a defined way for one program to ask another program for data or actionsUnit: AI from zero
- artifact
- a file a process produces, such as a built app package or a full tool outputUnit: Harnesses & context engineering
- asynchronous
- work that doesn't make you wait: you send it and the answer comes later; synchronous means you wait for itUnit: Local inference & serving
- attacker
- a person deliberately trying to break into or misuse a systemUnit: Prompt injection & agent security
- audit log
- a permanent record of who did what and when, kept so it can be checked laterUnit: Tool use & the agent loop
- auth
- short for authentication and authorisation: checking who someone is and what they may doUnit: Agent memory & MCP
- await
- waits for an asynchronous step to finish before the next line runsUnit: Evals & guardrails
B
- backfill
- filling in past or missing data in bulk, after the factUnit: Routing, cost & local vs API
- backup
- a copy of data kept so it can be restored after loss or damageUnit: Routing, cost & local vs API
- batching
- a server running several users' requests together in one pass on the GPUUnit: AI from zero
- benchmark
- a standard test or reference number used to compare againstUnit: Prompting & structured output
- billing
- working out and sending what each customer owesUnit: Routing, cost & local vs API
- branch
- to take a different path in code depending on a conditionUnit: Prompting & structured output
- browser
- the app used to open websites, like Chrome, Safari or FirefoxUnit: Tool use & the agent loop
C
- cache
- a fast store of recent results, so repeated requests don't redo the workUnit: Production agents
- CC
- to copy someone on an email so they get it tooUnit: Agent memory & MCP
- client code
- the program that calls a service; the service it calls is the serverUnit: Local inference & serving
- cluster
- a group of machines or items that work or sit togetherUnit: Harnesses & context engineering
- commit
- a saved snapshot of code changes in Git; to commit code is to save it into the project historyUnit: AI from zero
- constraint
- a limit or rule the solution has to respectUnit: AI from zero
- consumer
- an ordinary member of the public using a product, as opposed to a businessUnit: Local inference & serving
- contamination
- test items leaking into a model's training dataUnit: Eval design
- corpus
- a large collection of text; the plural is corporaUnit: AI from zero
- CPU
- the main processor chip that runs a computer's programsUnit: Local inference & serving
D
- database
- an organised store of data that programs can save to and look things up inUnit: Embeddings & RAG
- delay code
- the standard two-digit IATA code airlines record for the cause of a delayUnit: Embeddings & RAG
- delegate
- to hand a piece of work to someone, or some agent, elseUnit: Multi-agent orchestration
- dependency
- something your work relies on, such as a library your code uses or a task that must finish firstUnit: Production agents
- deploy
- to put a new version of software live where users can reach itUnit: AI from zero
- deterministic
- always giving the same output for the same inputUnit: AI from zero
- diff
- a view showing exactly which lines changed between two versionsUnit: Reward hacking & specification gaming
E
- encode
- to store information in another form; Base64 encoding turns any data into plain letters and digitsUnit: AI from zero
- endpoint
- one specific address of an API that you send requests toUnit: Local inference & serving
- escalate
- to pass a problem up to someone with more ability or authorityUnit: Routing, cost & local vs API
- eval
- a repeatable test set that scores how well a model or agent does a taskUnit: AI from zero
- expect(...).toBe(...)
- a test assertion: the test fails unless the value equals the expected oneUnit: Reward hacking & specification gaming
- exploit
- to take advantage of a flaw or weaknessUnit: Reward hacking & specification gaming
F
- few-shot prompting
- putting a few worked examples in the prompt so the model copies the patternUnit: Prompting & structured output
- fine-tuning
- training an existing model a bit more on your own examples to change its behaviourUnit: Tokens & next-token prediction
- flaky
- unreliable: sometimes works and sometimes fails for no clear reasonUnit: Production agents
- framework
- a ready-made structure to build on: a code framework, or a step-by-step method for thinkingUnit: Reward hacking & specification gaming
G
- generalisation
- a learned behaviour carrying over to new situationsUnit: Reward hacking & specification gaming
- grading
- scoring an output automatically, for example with code checks or another model as judgeUnit: Evals & guardrails
- greedy decoding
- always picking the single most likely next word; this is what temperature 0 doesUnit: AI from zero
H
- hallucination
- a fluent answer that is false or made upUnit: AI from zero
- handler
- the code or component that deals with a particular kind of request or inputUnit: Production agents
- HTTP
- the protocol browsers and APIs use to send requests and replies; HTTPS is the encrypted versionUnit: Local inference & serving
I
- incident
- an event that disrupts a service or puts data at risk and needs a responseUnit: AI from zero
- index
- a sorted lookup structure that lets a database or search engine find items without reading everythingUnit: Embeddings & RAG
- inference
- running a trained model to get answers, as opposed to training itUnit: Local inference & serving
- inject
- to insert something into a running system from outsideUnit: Fine-tuning & adaptation
- instruction tuning
- extra training on example questions and good answers so the model follows instructionsUnit: Tokens & next-token prediction
- invoice
- a bill sent to a customer listing what they oweUnit: Production agents
- irreversible
- cannot be undoneUnit: Evals & guardrails
L
- library
- ready-made code you add to your project instead of writing it yourselfUnit: Tool use & the agent loop
- log
- a timestamped record a program writes about what it didUnit: Evals & guardrails
M
- Markdown
- a simple text format where symbols become formatting, e.g.  shows an imageUnit: Agent memory & MCP
- merge
- to combine two things into one, like joining a code branch back into mainUnit: AI from zero
- monitoring
- continuously watching a system's health using the data (telemetry) it reportsUnit: Evals & guardrails
N
- next-token prediction
- the model writes by repeatedly guessing the most likely next piece of textUnit: AI from zero
- NVIDIA
- the company that makes most of the GPUs used to run AI modelsUnit: Local inference & serving
O
- outage
- a period when a service is down or not workingUnit: Routing, cost & local vs API
- overfitting
- fitting the test or training examples so closely that real-world results don't improveUnit: Prompting & structured output
P
- parse
- to read text and turn it into structured data a program can useUnit: AI from zero
- pipeline
- a chain of automated steps where each step's output feeds the nextUnit: Prompting & structured output
- Postgres (PostgreSQL)
- a popular open-source relational databaseUnit: Embeddings & RAG
- pre-training
- the first, biggest training stage: predicting the next token across a huge amount of textUnit: Tokens & next-token prediction
- preference tuning
- training that nudges the model toward answers people ranked higherUnit: Tokens & next-token prediction
- prefix
- the first part of a text or keyUnit: Production agents
- production
- the live system real users use, as opposed to test copiesUnit: Local inference & serving
- proxy
- a stand-in: a service that fetches things on your behalf, or a number that stands in for the real goalUnit: Agent memory & MCP
- pruning
- removing parts of a model to make it smaller and fasterUnit: Local inference & serving
- Pydantic
- a Python library that checks data against a defined schemaUnit: AI from zero
Q
- quarantine
- set aside or isolate something so it can't affect the rest until it is checkedUnit: Agent memory & MCP
- query
- a request for information written in a search or database languageUnit: Embeddings & RAG
R
- randomness
- chance: results that can differ each timeUnit: AI from zero
- README
- the introduction file at the top of a code project that explains what it is and how to use itUnit: Prompt injection & agent security
- red-teaming
- testing a system by attacking it on purpose to find weak spotsUnit: Prompt injection & agent security
- replay
- re-running recorded events or history to rebuild the same stateUnit: Production agents
- required
- the JSON Schema list of properties that must be presentUnit: Tool use & the agent loop
- retry
- try the same action again after it failsUnit: Prompting & structured output
- retry with backoff
- try again after a failure, waiting longer each timeUnit: Production agents
- RL (reinforcement learning)
- training a model by rewarding outputs a grader scores well, so it learns to get higher scoresUnit: Reward hacking & specification gaming
- rollout
- releasing a change to users, often step by stepUnit: Evals & guardrails
- roster
- the schedule of which staff work which shifts or flightsUnit: Prompting & structured output
- row
- one record in a database table, such as one customer or one flightUnit: Embeddings & RAG
- runtime
- the moment a program is actually running, or the software that runs itUnit: Fine-tuning & adaptation
S
- sampling
- picking one option at random, weighted by probability; for traces, keeping only some of themUnit: AI from zero
- sandbox
- a locked-down space where code can run without reaching anything importantUnit: Production agents
- SE (standard error)
- how much a measured score would wobble by chance; about ±2 SE gives a 95% rangeUnit: Agent evals & statistics
- server
- a computer that runs programs and answers requests from other machinesUnit: AI from zero
- SLA
- service level agreement: a contract promise about a service's level, with penalties if missedUnit: Local inference & serving
- snapshot
- a frozen copy of data at one momentUnit: Reward hacking & specification gaming
- stale
- out of date; freshness is how recent data isUnit: Embeddings & RAG
- Streamable HTTP
- MCP's network transport: messages travel over ordinary HTTP, with streaming for long repliesUnit: Agent memory & MCP
T
- table
- data laid out in rows and columnsUnit: Checkpoint: Core
- temperature
- a setting for how random the model's next-word pick is; 0 means always take the most likely oneUnit: AI from zero
- test suite
- a set of automatic checks run togetherUnit: Reward hacking & specification gaming
- timeout
- a time limit after which a waiting request gives upUnit: Production agents
- timestamp
- the recorded date and time something happenedUnit: Harnesses & context engineering
- tokenizer
- the part that splits text into a model's tokensUnit: Tokens & next-token prediction
- top-k
- keep only the k best-scoring resultsUnit: Embeddings & RAG
- triage
- sorting incoming items by urgency to decide what to handle firstUnit: Production agents
U
- UI (user interface)
- the screens, buttons and text a person uses to work an appUnit: Tool use & the agent loop
- unified memory
- one memory pool shared by CPU and GPU, as on a DGX SparkUnit: Local inference & serving
- unit test
- a small, fast test of one piece of code; end-to-end tests run the whole appUnit: Evals & guardrails
- utilisation
- how much of the available capacity is actually in useUnit: Routing, cost & local vs API
V
- validate
- check that data has the expected shape and values before using itUnit: AI from zero
Your first AI & LLMs → Agents lesson is five minutes away.
Not in the stores yet. Leave your email and we will tell you when it launches.