PART 2 - Taxonomy: What Kind of Agent Are You Building?
The 5 Levels - From Brain to Self-Evolving System
The Intuition
Think of hiring: you wouldn't hire a contractor without knowing the job scope. Same here - pick the right level of agent for the complexity of the task.
Level 0: Core Reasoning System (The Encyclopedia)
- What: LLM alone, no tools, no memory
- What it can do: Answer from training data
- Production use: FAQ chatbot on known, static content
- Limit: Knowledge cutoff. Can't act. Can't look things up.
Level 1: The Connected Problem-Solver
- What: LLM + external tools (Search, Calculator, DB queries)
- What it can do: Answer about real-time events, current data
- Production use: Customer support agent that can look up order status
- Technique: Function calling - the LLM decides which tool to use, ADK executes it
Level 2: The Strategic Problem-Solver
- What: LLM + tools + planning (multi-step, uses output of step N as input to step N+1)
- The Key Concept: Context Engineering - the output of one step is shaped into precise input for the next
- Production example (coffee shop case):
Step 1: maps_tool("Mountain View", "SF") → "Millbrae"
Step 2: search_tool("good coffee in Millbrae") <-- uses output from step 1- Production use: Research agent, code review pipeline, document analysis workflows
Level 3: The Collaborative Multi-Agent System
- What: A "Project Manager" agent + specialized sub-agents
- Why not one giant agent: Specialization. Each agent has a focused system prompt, relevant tools, and smaller context → cheaper, faster, more reliable
- Production example - Product Launch:
OrchestratorAgent --> MarketingAgent (writes press release)
--> WebDevAgent (builds landing page)
--> DataAgent (pulls competitor analysis)- Communication: Agents share state via
session.state(Google ADK) or A2A protocol
Level 4: The Self-Evolving System
- What: System creates new tools or new agents on the fly when it detects a capability gap
- Production use: Very experimental. Research/internal automation. Not for customer-facing production yet.
- The key skill: Agent identifies "I don't have a tool for this" and generates one dynamically
Session, State, and Memory
Understanding how agents track context across a conversation is foundational:
| Concept | What it is |
|---|---|
| Session | The current conversation thread |
State (session.state) | Key-value data within the current session - the agent's scratchpad |
| Memory | Searchable, cross-session information - persists beyond a single conversation |
ADK provides separate service implementations for each: SessionService manages session objects, MemoryService manages long-term memory.
State: The Session's Scratchpad
session.state is a dictionary of key-value pairs the agent reads and writes during a conversation. It holds everything the agent needs to recall to make the current conversation effective - user preferences set earlier, intermediate results, decisions already made.
# Writing to state
session.state["user_name"] = "Sachin"
session.state["preferred_currency"] = "EUR"
# Reading from state in the next turn
name = session.state.get("user_name")State is scoped to the session. When the session ends, state is gone - unless explicitly promoted to long-term Memory.
The A2A Protocol
The Agent2Agent (A2A) Protocol is an open standard for communication between independent AI agent systems. It enables agents to:
- Discover other agents' capabilities (via Agent Cards)
- Manage collaborative tasks asynchronously across agent boundaries
- Securely exchange information without exposing internal state, memory, or tools
A2A is the external communication layer - where session.state handles intra-agent state within a single runtime, A2A handles inter-agent coordination across separate systems.
The Production Trap
WARNING
Building at Level 3 or 4 before you've validated Level 1. Most production failures come from over-engineering the agent architecture before the individual agent is proven to work reliably on its own task.
Start at the lowest level that solves the problem. Add complexity only when you hit a real limit.
Recall Hook
Encyclopedia → Expert → Strategist → Manager → Self-Growing Org - pick your level before you write a line of code.
Sources
- Google ADK Whitepaper: Introduction to Agents
- Google ADK: Session, State, and Memory
- A2A Protocol Specification
Know a production example of each level? Add it here - real examples make this more useful.