| A2A |
Agent-to-Agent protocol. Google’s protocol for inter-agent communication. |
| AAIF |
Agentic AI Foundation. Linux Foundation project for agent infrastructure standards. |
| Agent Loop |
The core pattern of observe → think → decide → act → evaluate that all agents follow. |
| AGENTS.md |
Standard markdown file that tells AI agents how to work with a project. |
| Agentic Search |
Search where the agent decides what to search for, evaluates results, and iterates. Contrast with RAG. |
| Capability Token |
A JWT-like token encoding an agent’s permissions, time-bound and scope-limited. |
| Circuit Breaker |
Pattern that stops an agent from repeatedly failing by “opening” after N failures. |
| Conductor Model |
Engineering paradigm where humans orchestrate agents rather than writing code directly. |
| Context Engineering |
The discipline of optimizing what information goes into a model’s context window. |
| Context Pollution |
When redundant or irrelevant information in the context window degrades output quality. |
| Context Window |
The total amount of text (in tokens) a language model can process in a single request. |
| Defense in Depth |
Security approach using multiple layers of protection, each catching what others miss. |
| Delegation Chain |
When agents delegate tasks to other agents, each with a subset of the parent’s permissions. |
| Distill |
Open-source context engineering tool for deduplication and compression. |
| Episodic Memory |
Agent memory of past sessions, stored as summaries in a vector database. |
| Exfiltration |
Unauthorized extraction of data from a system, often through normal agent capabilities. |
| Hallucination |
When a language model generates incorrect or fabricated information. |
| Handoff |
Pattern where one agent transfers a task to another specialist agent. |
| Human-in-the-Loop |
Design pattern requiring human approval for certain agent actions. |
| Indirect Injection |
Prompt injection via data the agent processes (webpages, documents, tool responses). |
| Landlock |
Linux security module for restricting filesystem access at the kernel level. |
| MCP |
Model Context Protocol. Standard protocol for connecting AI agents to external tools. |
| Meta-MCP |
Pattern for compressing many MCP tool definitions into a few meta-tools. |
| MMR |
Maximal Marginal Relevance. Algorithm balancing relevance and diversity in retrieval. |
| OpenFGA |
Open-source Zanzibar implementation for fine-grained authorization. CNCF Incubating. |
| Prompt Injection |
Attack where malicious instructions are embedded in data the LLM processes. |
| RAG |
Retrieval-Augmented Generation. Pattern for giving LLMs access to external knowledge. |
| ReAct |
Reason + Act. Agent pattern with explicit reasoning before each action. |
| ReBAC |
Relationship-Based Access Control. Authorization based on relationships between entities. |
| seccomp |
Secure Computing Mode. Linux kernel feature for restricting system calls. |
| SKILL.md |
Markdown file with YAML frontmatter that encodes reusable procedural knowledge for AI agents. Installed via the skills.sh ecosystem. |
| Semantic Memory |
Long-term agent memory stored as a knowledge graph. |
| Session Memory |
Short-term agent memory within a single conversation/session. |
| Token |
The basic unit of text processing for language models. ~4 characters in English. |
| Token Budget |
Deliberate allocation of context window capacity across different purposes. |
| Tool Poisoning |
Attack where malicious instructions are embedded in tool/API responses. |
| Two-Layer Review |
Review process: Layer 1 (automated checks) + Layer 2 (human judgment). |
| Backpressure |
Automated feedback loops that catch agent errors before they reach human review. |
| Cascade Routing |
Model routing pattern that tries cheap models first and escalates to expensive ones only if quality is insufficient. |
| Ephemeral Environment |
A fresh, isolated environment created for each agent session and destroyed when complete. |
| Fleet-Scale Parallelism |
Running hundreds of identical agents against different targets simultaneously. |
| Golden Dataset |
A curated set of tasks with known-good outputs used for agent evaluation. |
| Kill Switch |
A mechanism to immediately terminate an agent session when it exhibits anomalous behavior. |
| LLM-as-Judge |
Using a language model to evaluate the quality of another model’s output. |
| Model Routing |
Sending different tasks to different models based on task characteristics and cost constraints. |
| Stall Detection |
Detecting when an agent is making the same tool call repeatedly without progress. |
| Zanzibar |
Google’s global authorization system. Basis for OpenFGA and similar systems. |