The paper outlines principles for designing successful AI agents, emphasizing simplicity, iterative complexity addition, and a deep understanding of underlying mechanisms. It advocates for starting with basic LLM API calls and gradually introducing frameworks or more complex patterns only when necessary, asserting that successful implementations tend to leverage simple, compoundable patterns rather than relying on overly complex frameworks.

A core distinction is made between Workflows and Agents:

• Workflows involve LLMs and tools operating along pre-defined, fixed code paths. This approach is preferred when predictability and consistency are paramount.
• Agents empower the LLM to dynamically manage its processes and tool usage, making autonomous decisions. This is suitable for scenarios requiring high flexibility and model-driven decision-making.

The paper suggests a progressive approach: begin with simple LLM calls or search functionalities, and if insufficient, incrementally introduce Workflows or Agents. When adopting frameworks like LangGraph, Bedrock, Rivet, or Vellum, a thorough understanding of their internal workings is crucial to avoid issues arising from abstraction.

Several practical workflow and agent patterns are detailed:

• Augmented LLM: Involves integrating built-in extensions such as search capabilities, tool connections, and memory. This pattern necessitates meticulous interface design and documentation.
• Prompt Chaining: Decomposes a complex task into a sequence of multiple LLM calls. Each call processes the output of the previous one, enhancing clarity and accuracy. For instance, generating marketing copy can be chained with a translation step, or document drafting with subsequent review and finalization steps.
• Routing: Classifies incoming inputs and directs them to appropriate processing paths or specific tools. Examples include routing customer inquiries based on type or channeling complex questions to more capable, high-performance models.
• Parallelization:
  • Sectioning: Divides a task into multiple sub-tasks that can be processed concurrently.
  • Voting: Executes the same task multiple times and selects the optimal result, a technique useful for tasks like code vulnerability review or automated LLM evaluation.
• Orchestrator-Workers: A master agent orchestrates and distributes sub-tasks to multiple worker agents, then integrates their outputs. This pattern is applicable for managing complex coding projects by distributing specific parts in real-time or for consolidating data from various sources.
• Evaluator-Optimizer: An iterative improvement loop where one LLM generates a response, and another LLM evaluates this response, providing feedback for subsequent refinements. This can be used for tasks like iteratively improving translation quality or synthesizing complex information.

For tool prompt engineering, the paper recommends designing tools with formats easily consumable by LLMs, allocating sufficient tokens, and providing clear tool descriptions including usage examples, edge cases, input format requirements, and distinct boundaries with other tools. Rigorous testing of the model's actual utilization of tools (e.g., via workbenches) and implementing "poka-yoke" (mistake-proofing) designs are also emphasized.

The overarching principles reiterate simplicity, transparency in agent planning, robust documentation and testing of tools/interfaces, and minimizing abstraction when using frameworks to maintain direct control.