2025_AI智能体构建实用指南（英）_33页_7mb

Large language model agents represent a new category of automated systems with advanced capabilities in reasoning, multimodality, and tool use.

What defines an agent?

• An agent autonomously executes workflows to achieve user goals.
• Core components: LLM for reasoning/decision-making, external tools for actions/data retrieval, clear instructions/guardrails.
• Key characteristics: Recognizes workflow completion, corrects actions, handles failures, dynamically selects tools within defined boundaries.

When (and when not) to build an agent?
Agents excel where traditional automation fails:

• Complex workflows with nuanced judgment, exceptions, or dynamic decision-making (e.g., refund approval context).
• Maintaining unwieldy systems with extensive brittle rules (vendor security reviews).
• Handling unstructured data and natural language interpretation (e.g., processing insurance claims, conversational interaction).

Best Practices: Agent Design Foundations

1. Model Selection:
   • Start with the best model for each task to establish performance.
   • Establish a performance baseline before optimizing.
   • Replace larger models with smaller ones where feasible for cost/latency benefits after validation.
2. Tool Definition:
   • Purpose: Access data/information or act on systems.
   • Key Types: Data Retrieval, Actions, Orchestration (agents themselves).
   • Equipment agents with reusable, documented, tested tools. Standardized definition facilitates flexible agent-tool relationships.
3. Instruction Configuration:
   • Based on existing procedures/scripts/policies.
   • Use clear, step-by-step prompts. Each prompt step should correspond to a specific action or output.
   • Capture edge cases and define handling procedures. Annotate decisions and handle deviations appropriately.
4. Orchestration Patterns:
   • Single-Agent: Simpler initially; start loops, use flexible prompts/variables; add tools incrementally. (Use when: many tasks, clearer instructions than complex prompts, manageable tools).
   • Multi-Agent:
     • Manager: Central agent coordinates specialized agents via tool calls (use model: efficiently manage complex control).
     • Decentralized Handoff: Peers hand off tasks based on specialization. (use model: direct control transfer across peers).
   • Graph representation helps visualize multi-agent systems.
5. Guardrails:
   • Critical for managing risks (data privacy, reputational).
   • Think multi-layered defense: Input safety/relevance, PII protection, Hallucination detection, Tool risk ratings, Rules-based validations, Human-in-the-loop for interventions.
   • Strategically add guardrails: Focus on data/brand safety, based on real-world failures, balance security and user experience.
   • Human intervention triggers: Failure thresholds (too many failures), High-risk actions (critical errors/exploitations). Allows smooth handoff to humans/users.

Conclusion & Getting Started

• Agents automate complex workflows intelligently.
• Build reliable agents by combining strong models, clear instructions, suitable tools, and appropriate orchestration.
• Start small, build iteratively, validate, and scale.
• Agents offer end-to-end execution capabilities suitable for complex, ambiguous, or context-dependent tasks.