FREE RESOURCE · FROM THE "AGENTIC" REEL
Agentic AI concepts: the 9 building blocks every builder should know, with a plain example for each
By Hammad Yousuf · 1 November 2026
An LLM alone only answers; memory, tools, evals and coordination turn it into a system that does work.
In four lines
- An LLM alone only answers; memory, tools, evals and coordination turn it into a system that does work.
- MCP connects an agent to tools and data; A2A connects agents to each other.
- Skills and RAG both bring in knowledge only when the task needs it.
- Evals tell you whether a change made the agent better or worse.
One agent, all 9 blocks
To keep the 9 blocks concrete, every one below is shown on the same small agent: one that checks supplier invoices before they are paid. It reads each invoice, matches it to the purchase order, flags anything odd and files the result.
Block 01
Memory & State
what the agent keeps between turns, and what it saves for later.
On the invoice agent
While it works through one invoice it keeps the purchase order it found (state). After it finishes, it saves that supplier's usual prices (memory), so next month a 30% jump stands out.
- You need it when
- The task spans more than one turn, or the next run should learn from this one.
- You can skip it when
- A one-shot job with everything it needs in the prompt.
Block 02
Orchestration
one planner splits the goal and hands each task to a worker.
On the invoice agent
A planner takes "check this week's invoices", splits it into one job per invoice and hands each to a worker, then collects the flags into one report.
- You need it when
- The goal breaks into parts that can run separately.
- You can skip it when
- A single step with one obvious order of work.
Block 03
RAG
search your own docs first, then answer from what it found.
On the invoice agent
Before judging a line item, it searches your contracts folder for that supplier's agreed rates and answers from the page it found, not from guesswork.
- You need it when
- The answer depends on documents the model has never seen.
- You can skip it when
- Everything it needs already fits in the prompt.
Official reference: RAG paper (Lewis et al., 2020) ↗
Block 04
Harness
the code around the model that feeds it context, runs its tools and holds the limits.
On the invoice agent
The harness hands the model the invoice text and the matching order, runs the lookup tools on its behalf, and stops the run after a set number of steps so a stuck lookup cannot loop forever.
- You need it when
- Always. A model without a harness can only reply; it cannot act.
- You can skip it when
- Never, though a first version can be a few dozen lines.
Block 05
Evals
fixed test cases plus a grader, so you know if a change helped or hurt.
On the invoice agent
Keep 20 past invoices with the known right answer (pay, hold, reject). Every time you change the prompt or the model, run all 20 and count how many it still gets right.
- You need it when
- Before you change anything in an agent that is already in use.
- You can skip it when
- You cannot skip it for long; without evals every change is a guess.
Block 06
MCP
an open protocol: one connector gives any MCP app its tools, resources and prompts.
On the invoice agent
Your accounting system exposes one MCP server with a tool (look up a purchase order), a resource (the supplier list) and a prompt (the invoice-check template). Any MCP client can use all three.
- You need it when
- You want the same tools and data available to more than one agent or app.
- You can skip it when
- One agent, one tool, and you control both ends.
Official reference: Model Context Protocol (MCP) specification ↗
Block 07
Skills
a SKILL.md folder the agent loads only when a task needs it.
On the invoice agent
A folder called invoice-check with a SKILL.md that holds your approval rules and the CSV format you file results in. Only its description sits in context until an invoice shows up.
- You need it when
- A procedure you repeat, with rules you do not want pasted into every prompt.
- You can skip it when
- A one-off task you will not run again.
Official reference: Agent Skills in Claude Code ↗
Block 08
A2A
an open protocol for agents to find each other and hand off tasks.
On the invoice agent
When an invoice fails the check, the invoice agent hands the case to a separate payments agent, which publishes an Agent Card describing what it can do and where to reach it.
- You need it when
- Two or more agents, often built by different teams, must pass work between them.
- You can skip it when
- Everything runs inside one agent, or one team's code can call the other directly.
Official reference: A2A protocol specification ↗
Block 09
Multi-Agent
focused agents, one role each, working toward one goal.
On the invoice agent
One agent reads invoices, one matches purchase orders, one writes the weekly summary. Each has one role and a small set of tools, and together they close the week's books.
- You need it when
- One agent's prompt and tool list have grown too large to reason about.
- You can skip it when
- A single focused agent still does the job well; add agents only when it stops.
Map your own agent onto the 9 blocks
Replace [JOB] with what your agent should do and paste it into Claude or any assistant.
Map an agent idea onto the 9 blocks
I want to build an AI agent that [JOB]. For each of these 9 blocks, tell me in one or two lines whether I need it and what it would look like for this job: memory & state, orchestration, RAG, harness, evals, MCP, skills, A2A, multi-agent. Then list the 3 blocks I should build first and the simplest version of each.Common questions
What is the difference between MCP and A2A?
MCP connects an agent to tools and data: an MCP server exposes tools, resources and prompts to any MCP client. A2A connects agents to each other: an agent publishes an Agent Card describing its identity, capabilities, skills and endpoint, and other agents hand it tasks that return artifacts.
What is the difference between RAG and skills?
Both bring in knowledge only when the task needs it. RAG retrieves documents at run time and adds them to the model input before it generates. A skill is a folder with a SKILL.md whose description sits in context until the skill is used, then its full instructions load.
Which of the 9 blocks should I build first?
The harness, because a model without one can only reply. Then evals, so every later change can be measured. The prompt on this page asks the model to pick the first three for your specific job.
Is a multi-agent system always better than one agent?
No. Multi-agent means focused agents with one role each working toward one goal. It helps once one agent's prompt and tool list have grown too large to reason about; before that, one agent is simpler to test and debug.
Where do the definitions on this page come from?
The one-line definitions are the ones in the reel. MCP, A2A, Agent Skills and RAG were checked against the official specifications and the original paper, linked at the bottom of the page.
Official specifications
Sources checked
- MCP servers expose three primitives: tools (model-controlled functions), resources (context data) and prompts (user-chosen templates); MCP is client-server. Source ↗
- An A2A Agent Card is a JSON document describing an agent's identity, capabilities, skills, service endpoint and auth; a task is the unit of work and returns its outputs as artifacts. Source ↗
- A skill is a folder with a SKILL.md (name, description); only descriptions load into context until the skill is used; Agent Skills is an open standard that works across tools. Source ↗
- RAG (retrieval-augmented generation): retrieve documents, add them to the model input, then generate (Lewis et al., 2020). Source ↗
About the video: the reel has no presenter; it is an animated board. The small headshot in it is an AI-generated image of Hammad Yousuf.