agent-foundations

name	agent-foundations
description	core cognitive architectures for Gemini agents. Use this to implement reasoning loops (ReAct, Chain-of-Thought) and understand the orchestration layer.

Gemini Agent Foundations

Goal

Design the "Orchestration Layer"—the cognitive loop that enables an agent to plan, execute, and adapt its actions to achieve a goal.

Cognitive Architectures

1. The ReAct Loop (Reason + Act)

• Concept: A cyclic process where the model alternates between internal reasoning and external action.
• The Loop:
  1. Thought: The agent analyzes the user's request and plans the next step.
  2. Action: The agent selects a tool (e.g., Flights) to execute.
  3. Action Input: The agent generates the specific parameters for that tool.
  4. Observation: The agent receives the output from the tool.
  5. Repeat: The loop continues until the agent determines it has enough info to answer.
• Prompt Structure:

  Question: [User Input]
  Thought: I need to check the flight status first.
  Action: check_flight_status
  Action Input: {"flight_number": "UA123"}
  Observation: [Tool Output]
  Thought: The flight is delayed. I should check connecting flights.
  ...
  Final Answer: Your flight is delayed.
  

2. Chain-of-Thought (CoT)

• Concept: A linear reasoning path best suited for complex logic or math problems where no external tools are needed. It enables reasoning capabilities through intermediate steps.
• Usage: Use when the task requires intermediate reasoning steps but is self-contained within the model's training data.

3. Tree-of-Thoughts (ToT)

• Concept: A branching reasoning strategy for exploration or strategic lookahead tasks. It generalizes CoT and allows the model to explore various thought chains.
• Usage: Best for scenarios with multiple potential solutions, allowing the agent to backtrack and explore different "thought branches."

Agent vs. Model

• Model: Knowledge is limited to training data; single inference based on user query; no native tool implementation.
• Agent: Knowledge extended through tools; managed session history for multi-turn inference; native cognitive architecture using reasoning frameworks.