ReAct+ 🔴
Need: The task requires iterative action — the model reasons, acts, observes the result, and repeats.
Explanation: Reasoning + Acting in a loop. Used natively in AI agents (tool-use, browsing, coding agents).
Prompt format:
“Thought: [reasoning]. Action: [tool]. Observation: [result]. … Repeat … Final Answer:“
QA usage example
“Thought: Checking the endpoint. Action: GET /health. Observation: 200 OK. Thought: Checking auth. Action: POST /login without token. Observation: 401. Final Answer: auth works correctly.”
When to use ReAct+?
In practice you don’t write ReAct+ manually — AI agents (Claude Code, Codex, Copilot) use this pattern natively. It is worth knowing so you understand HOW an agent thinks and why it takes each next step.
Tree of Thought (ToT) 🔴
Need: The problem has multiple paths — you want exploration of alternatives instead of a single route.
Explanation: An extension of CoT — the model generates multiple reasoning branches, evaluates them, and selects the best one. Simulates brainstorming.
Prompt format:
“Consider: [problem]. Generate 3 approaches. For each: pros, cons, risk. Choose the best.”
QA usage example
“How to organize E2E tests in microservices (12 services)? A: central repo. B: tests per service. C: contract testing + minimal E2E. Evaluate and choose.”
When to use ToT?
When you face an architectural or strategic decision — tool selection, test organization, migration approach. ToT forces comparison of options instead of accepting the first answer.
In the next post: P2P2 and Reflexion — planning before execution and self-correction.