agentic-patterns

Agentic Workflow Patterns

Complexity Decision Framework

• Start with the simplest solution: single LLM call with retrieval or in-context examples. Often sufficient.
• Use workflows (predefined code paths) for well-defined tasks where predictability and consistency matter.
• Use agents (LLM-directed tool use and planning) only for open-ended tasks where flexibility and model-driven decision-making are needed.
• Add complexity only when it demonstrably improves outcomes. Measure first; agentic systems trade latency and cost for task performance.

Workflow Patterns

Prompt Chaining

• Decompose a task into a sequence of steps; each LLM call consumes the previous step’s output.
• Add gates (programmatic checks) on intermediate steps to keep the process on track.
• Use when the task decomposes cleanly into fixed subtasks; trade latency for higher accuracy per step.

Routing

• Classify input and direct to specialized prompts, tools, or models.
• Use when distinct categories are better handled separately (e.g. easy questions → small model, hard → capable model; support types → different downstream flows).
• Classification can be LLM-based or a traditional classifier.

Evaluator-Optimizer

• One LLM generates a response; another evaluates and gives feedback in a loop.
• Use when you have clear evaluation criteria and iterative refinement adds measurable value (e.g. translation with critique, multi-round search with “need more?” decisions).

Agent Design Principles

• Simplicity: Keep agent design as simple as the task allows; avoid extra abstraction.
• Transparency: Expose the agent’s planning steps explicitly so behavior is interpretable.
• Agent-Computer Interface (ACI): Invest in tool documentation and ergonomics as seriously as in human-facing interfaces; unclear tools are a major source of failure.