Appendix B: Glossary

TermDefinition
A2AAgent-to-Agent protocol. Google’s protocol for inter-agent
communication.
AAIFAgentic AI Foundation. Linux Foundation project for agent
infrastructure standards.
Agent LoopThe core pattern of observe → think → decide → act → evaluate that
all agents follow.
AGENTS.mdStandard markdown file that tells AI agents how to work with a
project.
Agentic SearchSearch where the agent decides what to search for, evaluates results,
and iterates. Contrast with RAG.
Capability TokenA JWT-like token encoding an agent’s permissions, time-bound and
scope-limited.
Circuit BreakerPattern that stops an agent from repeatedly failing by “opening”
after N failures.
Conductor ModelEngineering paradigm where humans orchestrate agents rather than
writing code directly.
Context EngineeringThe discipline of optimizing what information goes into a model’s
context window.
Context PollutionWhen redundant or irrelevant information in the context window
degrades output quality.
Context WindowThe total amount of text (in tokens) a language model can process in
a single request.
Defense in DepthSecurity approach using multiple layers of protection, each catching
what others miss.
Delegation ChainWhen agents delegate tasks to other agents, each with a subset of the
parent’s permissions.
DistillOpen-source context engineering tool for deduplication and
compression.
Episodic MemoryAgent memory of past sessions, stored as summaries in a vector
database.
ExfiltrationUnauthorized extraction of data from a system, often through normal
agent capabilities.
HallucinationWhen a language model generates incorrect or fabricated
information.
HandoffPattern where one agent transfers a task to another specialist
agent.
Human-in-the-LoopDesign pattern requiring human approval for certain agent actions.
Indirect InjectionPrompt injection via data the agent processes (webpages, documents,
tool responses).
LandlockLinux security module for restricting filesystem access at the kernel
level.
MCPModel Context Protocol. Standard protocol for connecting AI agents to
external tools.
Meta-MCPPattern for compressing many MCP tool definitions into a few
meta-tools.
MMRMaximal Marginal Relevance. Algorithm balancing relevance and
diversity in retrieval.
OpenFGAOpen-source Zanzibar implementation for fine-grained authorization.
CNCF Incubating.
Prompt InjectionAttack where malicious instructions are embedded in data the LLM
processes.
RAGRetrieval-Augmented Generation. Pattern for giving LLMs access to
external knowledge.
ReActReason + Act. Agent pattern with explicit reasoning before each
action.
ReBACRelationship-Based Access Control. Authorization based on
relationships between entities.
seccompSecure Computing Mode. Linux kernel feature for restricting system
calls.
SKILL.mdMarkdown file with YAML frontmatter that encodes reusable procedural
knowledge for AI agents. Installed via the skills.sh ecosystem.
Semantic MemoryLong-term agent memory stored as a knowledge graph.
Session MemoryShort-term agent memory within a single conversation/session.
TokenThe basic unit of text processing for language models. ~4 characters
in English.
Token BudgetDeliberate allocation of context window capacity across different
purposes.
Tool PoisoningAttack where malicious instructions are embedded in tool/API
responses.
Two-Layer ReviewReview process: Layer 1 (automated checks) + Layer 2 (human
judgment).
BackpressureAutomated feedback loops that catch agent errors before they reach
human review.
Cascade RoutingModel routing pattern that tries cheap models first and escalates to
expensive ones only if quality is insufficient.
Ephemeral EnvironmentA fresh, isolated environment created for each agent session and
destroyed when complete.
Fleet-Scale ParallelismRunning hundreds of identical agents against different targets
simultaneously.
Golden DatasetA curated set of tasks with known-good outputs used for agent
evaluation.
Kill SwitchA mechanism to immediately terminate an agent session when it
exhibits anomalous behavior.
LLM-as-JudgeUsing a language model to evaluate the quality of another model’s
output.
Model RoutingSending different tasks to different models based on task
characteristics and cost constraints.
Stall DetectionDetecting when an agent is making the same tool call repeatedly
without progress.
ZanzibarGoogle’s global authorization system. Basis for OpenFGA and similar
systems.