A practical knowledge base for AI protocols, agents, core concepts, frameworks, and tools.
A voice agent is an AI that listens, speaks, and actually gets things done, chaining speech recognition, reasoning, tool calls, and speech synthesis into a real-time loop. What separates it from an old voice assistant is that it has no scripted decision tree; what separates it from a text agent is latency and interruption handling
Agent Plugins is an open packaging format released in August 2026. A plugin uses plugin.json to bundle Agent Skills and MCP server configuration into one distributable directory that compatible agent clients can discover and load.
An Agent Swarm is a system in which relatively independent AI agents explore in parallel, share discoveries, and coordinate results. It can expand search coverage and total work, but agents based on similar models may repeat errors, duplicate effort, or deadlock on complex dependencies.
Agentic Payments lets an AI agent initiate and complete a payment within user-defined boundaries for identity, amount, counterparty, purpose, and time. The point is not an unlimited AI credit card, but machine-enforceable authorization, risk controls, execution, and auditable receipts.
Agentic Resource Discovery (ARD) is a federated open specification that lets AI agents find, describe, and verify MCP servers, A2A agents, Skills, APIs, and other capabilities across organizations without hard-coding the entire tool catalog into context.
MCP Tasks is an official MCP extension for asynchronous long-running tool calls. A server returns a task handle first; the client can then check status, provide more input when supported, and retrieve the result after completion instead of holding one request open.
Loop Engineering means you stop writing prompts for an AI Agent one at a time and instead design a system that finds its own work, runs it, and checks the results. You move from being the person who asks to being the person who designs the loop, its goals, its acceptance criteria, and its fallbacks.
Context Compaction is what happens when a long session is about to fill the context window: the system swaps earlier history for a summary that keeps progress, conclusions, and next steps while discarding verbose tool output, so the task can keep running instead of being forced to restart.
Context Engineering is about deciding, within a finite context window, which information actually goes into the model each turn: system instructions, tool definitions, retrieved material, message history, external memory. What it cares about isn't wording but the selection and organization of information.
Agentic AI refers to the class of AI systems that don't just generate content but plan their own steps, call tools, adjust based on results, and see a task through. The dividing line with generative AI is whether the output is content or action.
Evals are tests for an AI system: give it a set of inputs, then apply grading logic to judge whether the output is good. They turn "it feels worse this week" into a reproducible, comparable score you can gate a pipeline on — the precondition for daring to change a prompt or swap a model.
MoE (Mixture of Experts) splits a model's feed-forward layers into many "experts" and has a router decide which few each token passes through. Total parameter count can be enormous while only a small fraction actually computes on any given step.
Showing 12 of 12 entries
Every entry grouped by topic — one click straight to the explanation.
