A practical knowledge base for AI protocols, agents, core concepts, frameworks, and tools.
Test-time Compute means spending more computation at the moment you ask, in exchange for accuracy: letting the model think longer, try several paths, and check its own work. It changes no parameters and requires no retraining — it's the mechanism behind 2026's reasoning models answering better by thinking longer.
A Subagent is an independent instance the main agent dispatches: its own context window, its own tool permissions, returning only a conclusion so the digging never pollutes the main session. The core problem it solves is who gets to spend the context.
A World Model is a model that can internally simulate how an environment changes in response to actions: give it a current state and an action, and it predicts what happens next. Language models predict the next word; world models predict the next state.
Agent Skills is an open format for packaging "how to do this" for an AI: a folder plus a SKILL.md spelling out the procedure, with scripts and reference material alongside. MCP lets an agent use tools; Skills teach it what to do with them.
Guardrails are the validation and control layers wrapped around a model's inputs, tool calls, and outputs: check before the model, validate after it, intercept before an action executes. They run outside the model's context, so "ignore previous instructions" can't route around them.
Prompt Injection is when attacker-crafted text gets a model to treat "data" as "instructions" and act on it, bypassing the rules it was given. OWASP ranks it as the top risk for LLM applications (LLM01), and it's especially dangerous in agents that can read web pages and files.
Harness Engineering can be understood as the engineering practice of "fitting an AI Agent with a controllable working environment." What it focuses on isn't writing yet another magic prompt, but using context, tools, permissions, feedback, testing, and auditing to make the AI's actions more stable and more checkable.
A knowledge graph is a way of organizing knowledge with "entities" and "relationships." Unlike ordinary documents, which just store passages of text, it connects the relationships among people, companies, products, places, and concepts into a queryable knowledge network.
AI alignment studies how to make an AI system's behavior consistent with human values, intentions, and interests. The problem sounds simple but is extremely har
Prompt Caching is an APIlevel optimization technique that lets an identical prompt prefix be processed only once across multiple requests, with subsequent calls
AI search is a way of searching that combines realtime web retrieval with the generative capabilities of large language models. After a user asks a question, th
Computer Use is a capability that lets an AI model directly operate a computer's graphical interface. It doesn't rely on API interfaces; instead, like a human,
Showing 12 of 12 entries
Every entry grouped by topic — one click straight to the explanation.
