An encyclopedia of agents & their technology
A field guide to
agentic AI.
Understand the ideas, tools, and protocols behind AI agents. Follow the connections. Go back to the source.
Explore the libraryThe starting collection
01 — 20Core ideas, connected through sources and related entries.
- 01Foundations
AI agent
A system that uses a model to choose actions, observe their results, and work toward a goal within defined limits.
- 02Capabilities
Tool use
The structured handoff between a model’s request and software that can retrieve information or take an action.
- 03Protocols
Model Context Protocol
An open protocol for connecting AI applications to tools, resources, and prompts through a shared interface.
- 04Foundations
Agent orchestration
How software coordinates model decisions, tools, state, and delegated work while retaining clear stopping and authority boundaries.
- 05Capabilities
Agent memory
How an agent stores and retrieves information across steps or sessions, and why persistence, relevance, and truth are separate questions.
- 06Capabilities
Retrieval-augmented generation
A way to generate responses with retrieved material, with distinct failure points in finding evidence, selecting context, and using it faithfully.
- 07Foundations
Agent evaluation
How to test an agent’s outcomes, interaction history, reliability, and limits without confusing a benchmark score with a deployment guarantee.
- 08Foundations
Prompt injection
When untrusted text attempts to redirect an AI system, crossing the boundary between material to process and instructions the system should follow.
- 09Capabilities
Agent planning
How an agent represents possible steps toward a goal, checks intermediate results, and revises a course of action when assumptions fail.
- 10Capabilities
Durable agent execution
How an agent workflow records progress and resumes after interruption, while distinguishing replayed results from repeated external actions.
- 11Protocols
Tool permissions
The boundaries that determine who may request a tool action, which resource it may affect, and where authorization is enforced.
- 12Capabilities
Agent sandboxing
How an execution environment restricts code and tools, and why filesystem, network, resource and application boundaries need separate examination.
- 13Foundations
Agent observability
How traces, metrics and event records help explain an agent run, while keeping operational evidence distinct from factual evaluation and audit authority.
- 14Foundations
Human oversight
How people can inspect, interrupt, correct and govern agent behavior, with explicit distinctions between action approval, later review and technical enforcement.
- 15Foundations
Context engineering
How an agent selects and maintains the information available for each model call, with separate boundaries for memory, retrieval and authority.
- 16Capabilities
Structured outputs
How schemas constrain the form of model responses while leaving factual correctness, supported features and execution authority as separate checks.
- 17Protocols
Agent identity and delegation
How software establishes an agent caller’s identity and represents delegated authority without confusing a name or persona with verified permission.
- 18Protocols
Agent-to-agent communication
How agents exchange messages, task state and artifacts across an interface, with separate checks for interoperability, authority and result quality.
- 19Foundations
Agent budgets and termination
How an agent bounds turns, requests, time and paid work, and reports incomplete or uncertain outcomes when a limit stops execution.
- 20Capabilities
Computer-use agents
How agents observe and act through browser or desktop interfaces, with separate checks for screen grounding, action effects, isolation and authority.