Glossary
AI, in
working
terms.
30 terms across 4 levels. Written for operators who use AI to ship work, not for the discourse.
30
terms
4
levels
By Level
Jump to alphabetical ↓Level 1
Beginner
8 terms
AgentAn AI system that takes a goal, plans steps, uses tools, and returns a result without being walked through every move.
4 min→
Context WindowHow much text the model can see at once. Bigger is better until cost catches up.
3 min→
HallucinationWhen the model produces a confident answer that is not grounded in reality. The single biggest reason agents need verification.
4 min→
InferenceRunning a trained model to produce an output. The thing you pay for every time the agent thinks.
3 min→
PersonaA configured agent with a specific voice, role, and toolset. Same engine, different operator.
3 min→
PromptThe input you give the model. Includes the system instructions, the conversation, and any retrieved context.
4 min→
SkillA packaged capability an agent can install and use. Code, prompts, and tool wiring bundled as one unit.
3 min→
TokenThe unit a model reads and writes. Roughly four characters in English. Pricing is per token.
2 min→
Level 2
Intermediate
8 terms
Chain of ThoughtPrompting the model to reason step by step before answering. Lower error rates on multi-step problems, higher token cost.
5 min→
EmbeddingA numeric fingerprint of a chunk of text. Lets you find related content by distance instead of keyword match.
5 min→
EvalsAutomated tests for agents and prompts. Run on a fixed dataset, score the output, catch regressions before users do.
6 min→
Fine-tuningTraining a base model on your own data so it specializes. Less common now that prompting and RAG cover most cases.
6 min→
MCPModel Context Protocol. Open standard for connecting models to tools, data, and prompts. Standardizes the agent-to-app handoff.
7 min→
RAGRetrieval-augmented generation. Pull relevant snippets from your data, then ask the model. Trades training cost for retrieval cost.
6 min→
System PromptTop-of-conversation instructions that set the model's role, constraints, and style. Sticks for the whole session.
5 min→
Tool UseWhen the model calls an external function (search, code, calendar) instead of just writing text back.
6 min→
Level 3
Advanced
8 terms
Agent LoopThe plan-act-observe cycle. Model proposes a step, runs it, reads the result, decides the next step.
7 min→
GuardrailsHard limits on what an agent can do. Tool allowlists, output filters, sandboxed execution, human checkpoints.
7 min→
Human in the LoopPausing the agent for human approval at chosen points. Common pattern for high-stakes actions like sending money.
5 min→
MemoryWhat the agent keeps between sessions. Long-term store, not the active context window. Lets it learn your patterns.
6 min→
Model RoutingSending each request to the cheapest model that can handle it. Big win on cost without dropping quality on hard tasks.
6 min→
Prompt InjectionAdversarial input that hijacks an agent's instructions through user data or fetched content. The XSS of the agent era.
8 min→
SandboxingRunning agent code in an isolated environment with no host access. Required for any agent that runs untrusted code.
6 min→
Vector StoreA database for embeddings. Returns the nearest matches by cosine or dot product, not by exact keyword.
8 min→
Level 4
Expert
6 terms
Agent HarnessThe runtime around the model. Manages tool calls, context, memory, retries, error handling, and observability.
8 min→
Base ModelThe foundation weights before instruction tuning, RLHF, or fine-tuning. The raw next-token predictor.
8 min→
Computer UseAgents that drive a real or virtual desktop. Mouse, keyboard, screenshots. Can use any app, including ones with no API.
9 min→
Mixture of ExpertsArchitecture that routes tokens to a sparse subset of specialist subnetworks. Larger total parameters, lower active cost per token.
9 min→
RLHFReinforcement learning from human feedback. The training stage that turns a base model into a helpful assistant.
9 min→
x402HTTP 402 Payment Required, revived as a per-call payment protocol for agents. Powers Agentic Market.
7 min→
Alphabetical
A
3 terms
AgentAn AI system that takes a goal, plans steps, uses tools, and returns a result without being walked through every move.
Beginner4 min→
Agent HarnessThe runtime around the model. Manages tool calls, context, memory, retries, error handling, and observability.
Expert8 min→
Agent LoopThe plan-act-observe cycle. Model proposes a step, runs it, reads the result, decides the next step.
Advanced7 min→
C
3 terms
Chain of ThoughtPrompting the model to reason step by step before answering. Lower error rates on multi-step problems, higher token cost.
Intermediate5 min→
Computer UseAgents that drive a real or virtual desktop. Mouse, keyboard, screenshots. Can use any app, including ones with no API.
Expert9 min→
Context WindowHow much text the model can see at once. Bigger is better until cost catches up.
Beginner3 min→
E
2 terms
H
2 terms
M
4 terms
MCPModel Context Protocol. Open standard for connecting models to tools, data, and prompts. Standardizes the agent-to-app handoff.
Intermediate7 min→
MemoryWhat the agent keeps between sessions. Long-term store, not the active context window. Lets it learn your patterns.
Advanced6 min→
Mixture of ExpertsArchitecture that routes tokens to a sparse subset of specialist subnetworks. Larger total parameters, lower active cost per token.
Expert9 min→
Model RoutingSending each request to the cheapest model that can handle it. Big win on cost without dropping quality on hard tasks.
Advanced6 min→
P
3 terms
PersonaA configured agent with a specific voice, role, and toolset. Same engine, different operator.
Beginner3 min→
PromptThe input you give the model. Includes the system instructions, the conversation, and any retrieved context.
Beginner4 min→
Prompt InjectionAdversarial input that hijacks an agent's instructions through user data or fetched content. The XSS of the agent era.
Advanced8 min→
R
2 terms
S
3 terms
SandboxingRunning agent code in an isolated environment with no host access. Required for any agent that runs untrusted code.
Advanced6 min→
SkillA packaged capability an agent can install and use. Code, prompts, and tool wiring bundled as one unit.
Beginner3 min→
System PromptTop-of-conversation instructions that set the model's role, constraints, and style. Sticks for the whole session.
Intermediate5 min→
T
2 terms