Interview: Agents and Tool Use

The model is given structured descriptions of available tools (name, description, parameter schema) in its context. Instead of a plain text response, it's trained/prompted to output a structured request to call a specific tool with specific arguments when appropriate; the application code executes that call and feeds the tool's result back into the conversation for the model to use in its next response.

ReAct interleaves reasoning traces ("thought") with actions (tool calls) and observations (tool results) in a loop: the model thinks about what it needs, takes an action, observes the result, and repeats until it has enough information to produce a final answer - combining CoT-style reasoning with the ability to gather real information rather than reasoning purely from parametric knowledge.

A single agent has one LLM loop that reasons and calls tools directly. Multi-agent systems split work across specialized agents (e.g. a planner, a researcher, a coder, a critic) that communicate and hand off subtasks - potentially improving quality on complex tasks through specialization and self-critique, at the cost of significantly more complexity, latency, and failure modes (agents miscommunicating, compounding errors).

• Infinite/unproductive loops - mitigate with max iteration limits and loop detection.
• Error compounding - an early wrong tool result or reasoning mistake propagates through later steps; mitigate with verification steps and the ability to backtrack.
• Over-triggering tool use (calling tools when not needed) or under-triggering (failing to call a needed tool) - mitigate with clear tool descriptions and few-shot examples of correct usage.
• Unsafe/irreversible actions - mitigate with human-in-the-loop confirmation for high-stakes actions (sending emails, deleting data, spending money) and least-privilege tool permissions.

• Short-term: current conversation in the context window.
• Long-term: vector store or DB of past sessions and facts.
• Working memory: scratchpad the agent writes to during multi-step tasks.
• Episodic: logs of prior tool calls and outcomes for reflection.

ReAct interleaves one thought and one action at a time. Plan-and-execute first generates a full plan, then executes steps (possibly with a cheaper executor model). Planning reduces wasted tool calls on complex tasks but can fail if the initial plan is wrong without replanning.