Agents White Paper
Google's comprehensive whitepaper distinguishing agents from plain models, explaining how tools, orchestration layers, and cognitive architectures combine to create capable AI agents.
Overview
Google’s Agents whitepaper provides a formal framework for understanding what makes an AI agent more than just an LLM. The core argument: a model alone is not an agent. An agent emerges from the combination of a model, tools, and an orchestration layer that manages the interaction loop.
Agent vs. Model
The paper draws a clear line:
- A model takes input and produces output — it has no state, no tools, and no ability to act on the world.
- An agent wraps a model in a loop that can observe the environment, plan actions, execute tools, and incorporate feedback.
This distinction matters because it clarifies where to invest engineering effort: not just in model quality, but in the orchestration and tooling layers.
The Orchestration Layer
The orchestration layer is the “brain” that manages the agent’s execution loop:
- Receive input (user query or environment observation)
- Plan the next action using the model’s reasoning capabilities
- Execute the chosen action (tool call, API request, code execution)
- Observe the result and update internal state
- Repeat until the task is complete or a termination condition is met
The paper discusses several orchestration strategies, including ReAct-style interleaving of reasoning and action, and plan-then-execute approaches where the full plan is generated upfront.
Tool Integration
Tools are categorized into:
- Extensions — first-party APIs tightly integrated with the agent
- Functions — developer-defined capabilities exposed to the model
- Data stores — retrieval systems (RAG) that provide dynamic context
The paper emphasizes that tool descriptions serve as a form of in-context learning — the model uses the tool’s name, description, and parameter schema to decide when and how to invoke it.
Cognitive Architectures
The whitepaper introduces the concept of cognitive architectures for agents — structured ways of combining reasoning, memory, and action:
- Reactive — simple stimulus-response, no planning
- Deliberative — explicit planning before acting
- Reflective — self-monitoring and strategy adjustment during execution
Key Takeaway
Building an effective agent requires thinking beyond the model. The orchestration layer and tool interface are equally important — and often more impactful to improve than the underlying model itself.