1. Harness
The harness is the loop that turns a language model into an agent. It runs context into the model, takes the tool call the model wants, runs it, feeds the result back in, and goes again. That loop is the wiring that connects the model to its skills, memory, prompts, APIs and browser. It is also the thing a lot of engineers have started obsessing over, on the view that a weak harness fails more often than a weak model.
2. Memory and state
Memory is how an agent holds on to information across steps and tasks instead of starting cold every time. In practice it splits into short-term memory for the job in front of it, long-term memory for what it learned before, a sense of its current state, and a list of tasks still open. Without it, an agent forgets what it was doing the moment the context runs out.
3. RAG
Retrieval-Augmented Generation is how an agent looks things up before it answers. The flow is simple: take the question, search for relevant data, hand that data to the model, then generate the answer. The point is to ground responses in real, external sources rather than whatever the model happens to remember, which is a large part of how you cut down on confident wrong answers.
4. MCP
The Model Context Protocol is a standard way to plug an AI system into outside tools and data. Think of it as a common connector, so every tool does not need its own custom integration. It has quietly become one of the most important pieces of plumbing in the whole space, which is why rival companies keep building standards on top of it.
5. Skills
Skills are reusable instructions, knowledge, scripts and workflows that an agent loads only when a task calls for them. The value is that you do not have to stuff everything into one enormous prompt. The agent reaches for the right skill when it needs it and ignores the rest, which keeps it both more capable and easier to manage.
6. Guardrails
Guardrails are the rules that decide what an agent can and cannot do. They cover permissions, sandboxing, when a human has to approve something, and logging every action it takes. The whole aim is useful autonomy without uncontrolled behaviour, which is exactly the line the big labs have been arguing over as they let agents act with less sign-off.
7. Evals
Evals are how you measure whether an agent is actually any good, rather than just impressive in a demo. A proper setup tests before launch, measures real outcomes, feeds that back into improvements, and keeps monitoring once it is live. It is the unglamorous part, and it is the part that tells you if the thing works.
8. A2A
Agent-to-agent communication lets separate agents talk to each other, pass tasks back and forth and share results. The payoff is specialisation. Instead of one agent trying to do everything, different agents can each handle the job they are best at.
9. Multi-agent systems
A multi-agent system puts one orchestrator in charge of several specialised agents. A typical chain might run a researcher, then a coder, then a reviewer, before handing back a finished result. It is the team-of-agents idea, with a manager on top keeping them pointed at the same goal.
The big picture
Stack them together and the shape is clear. A language model, plus memory, plus retrieval, plus tools, plus skills, plus guardrails, plus evals, plus a way for agents to talk plus coordinate, is what turns a basic AI model into a genuine agentic system. The reason the cheat sheet is worth saving is that none of these are abstract anymore. Every one of them has driven real news in the past month, from harness prompts to MCP standards to the fight over how tight an agent's guardrails should be. Learn the nine words, and the headlines stop being noise.