ReAct, Plan-and-Execute, Reflexion: turning LLMs into goal-driven actors.

Agent Patterns — Tools, Memory, and Loops

An LLM agent combines reasoning, tool use, and memory in a loop until a goal is met. Patterns like ReAct (think→act→observe), Plan-and-Execute (decompose first), and Reflexion (self-critique) provide reliable scaffolding around the model's autonomous decisions.

Interactive Pipeline Flow

IngestProcessServe

🎯 Goal — user task

User provides a goal: "find my last 3 invoices and email them to accounting".

Architecture Diagrams

ReAct loop: Thought → Action → Observation, repeated until Done.

Loading diagram...

Key Design Decisions

🛠️Tool Quality > Count

Each new tool grows the prompt and confuses the model. Curate to <10 tools per agent; merge similar ones.

🔁Cap the Loop

Always set max_iterations (e.g., 10). Runaway agents burn tokens. Add a "give up gracefully" exit.

💾Scratchpad Drift

Long ReAct chains exceed context. Summarize older steps; keep only the last few raw observations.

🎯Plan First for Complex

For multi-step goals, Plan-and-Execute beats ReAct: cheaper plan with mini model, execute with stronger model.

🔄Reflexion Helps

On failure, ask the LLM to critique its trajectory and retry. Lifts task success ~10-20% with minor cost.

📋Trace Every Step

Persist (thought, action, observation) tuples. Critical for debugging "why did it do that?" in production.

Discussion