agents

Agents Skill

Agent Architectures

ReAct (Reason + Act) — Default Pattern

Thought: What do I need to figure out?
Action: [tool_name] with [parameters]
Observation: [tool result]
Thought: What does this tell me? What's next?
Action: ...
Final Answer: [synthesized result]

Best for: Research, data gathering, multi-step Q&A

Plan-and-Execute

Step 1: Planner agent creates a full plan (list of steps)
Step 2: Executor agent runs each step in order
Step 3: Replanner reviews results and updates plan if needed

Best for: Complex tasks with many interdependent steps

Reflection / Self-Critique Loop

Generate output -> Critique output -> Revise -> Repeat until quality threshold met

Best for: Writing, code generation, analysis quality improvement

Key implementation detail: define explicit quality criteria before entering the loop. Without measurable criteria, reflection loops spin without converging. Common criteria: factual accuracy, completeness against a checklist, adherence to a style guide, or passing a test suite.

Tree of Thoughts (ToT)

Step 1: Generate N candidate next-steps (branches)
Step 2: Evaluate each branch with a scoring heuristic
Step 3: Expand the top-K branches
Step 4: Repeat until a solution branch reaches the goal

Best for: Problems with large search spaces — math proofs, puzzle solving, strategic planning. More expensive than ReAct (requires multiple LLM calls per step), so use only when single-path reasoning fails.

Language Agent Tree Search (LATS)

Step 1: Select a node using UCT (Upper Confidence bound for Trees)
Step 2: Expand by generating candidate actions
Step 3: Evaluate with environment feedback + LLM self-reflection
Step 4: Backpropagate scores up the tree
Step 5: Repeat until budget exhausted or solution found

Best for: Complex reasoning tasks where you want Monte Carlo Tree Search-style exploration combined with LLM reasoning. Produces higher-quality results than ReAct on hard problems at the cost of more LLM calls.