Agentic AI Notebook
0What is an AI Agent?1Why LLMs Need Agents2Anatomy of an Agent3Agent Lifecycle4Core Concepts5Agent Capabilities6Types of Agents7Agent Architectures8Agent Terminology9Current Agent Landscape10Planning11Reflection12Multi Tool13Self Correction14Build First AI Agent
Agent Foundations
Phase 4Module 3 of 15

Anatomy of an Agent

Treating agents as just a prompt misses tool routing, recovery, and observability — demos break in production.

Brain judges, senses perceive, hands act, notebook remembers, manager orchestrates policy.

Visual Workflows

Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.

100%
Loading diagram...

Scroll inside the frame to explore · use + / − to zoom up to 200%

Data Flow Through Layers

100%
Loading diagram...

Scroll inside the frame to explore · use + / − to zoom up to 200%

One iteration: senses feed brain, brain selects hands, hands update memory.

Key Takeaways

  • 1.Brain = LLM plus system prompt and planning strategy.
  • 2.Senses = inputs from user, files, webhooks.
  • 3.Hands = tool registry with schemas and permissions.
  • 4.Memory = working buffer plus vector long-term store.
  • 5.Nervous system = runtime, retries, routing, HITL.

Real Example

Scenario

Cursor-style coding agent: Brain = Claude with repo rules; Senses = open files + terminal stderr; Hands = `write_file`, `run_tests`, `grep`; Memory = repo index + session buffer; Runtime = 12-step cap, retry on test failure, LangSmith trace.

What you would do

Brain reads failing test output (Senses) and chooses `write_file` then `run_tests` (Hands). Memory stores the last failing assertion so the retry does not repeat the same edit. Runtime stops at 12 steps or escalates to the user. Monitor: tools per successful fix and invalid file paths suggested by the brain.

Practice Task

Pick a non-coding agent (support, research, or ops). Fill a 5-row table — Brain / Senses / Hands / Memory / Runtime — with one concrete item per row for that agent.

Code Walkthrough

Highlighted lines show where Anatomy of an Agent happens in the code.

Anatomy of an Agent
1from typing import TypedDict  # import dependencies2
3class AgentState(TypedDict):  # define a data structure or component4    messages: list          # Brain input (Senses)5    tool_results: list      # Observations from Hands6    repo_index: dict        # Long-term Memory7    step: int               # Runtime counter8    max_steps: int          # Runtime limit9
10state: AgentState = {11    "messages": [{"role": "user", "content": "Fix the failing auth test"}],12    "tool_results": [],13    "repo_index": {"files": ["auth.py", "test_auth.py"]},14    "step": 0,15    "max_steps": 12,16}17
18while state["step"] < state["max_steps"]:  # key line for Anatomy of an Agent19    # Brain: LLM picks tool from registry (Hands)20    action = "run_tests"  # e.g. from tool_calls21    observation = {"passed": False, "stderr": "AssertionError: 401"}22    state["tool_results"].append(observation)23    state["step"] += 1

Cheat Sheet

Quick recap

quick ref
  • •5 layers: brain senses hands memory runtime
  • •Each layer independently testable
  • •Tool registry = schemas + permissions
  • •Memory tiers: working short long episodic
  • •Runtime handles retries and HITL

Common Mistakes

  • ✕Skipping evaluation for Anatomy of an Agent before production
  • ✕No logging or tracing around anatomy of an agent steps
  • ✕Ignoring cost and latency implications