The ReAct loop

question -> [LLM] Thought + Action -> [Tool] Observation
            ^                                       |
            |_______________________________________|
                       up to max_steps

ReAct alternates Reasoning (Thought + Action) with Action (observation from a tool) until the model emits a Final Answer:. It is the simplest agent pattern that works on small open models — and the easiest to debug.

The loop, in code

for step in range(max_steps):
    prompt = system + question + scratchpad_of_prior_steps
    raw = llm.generate(prompt, stop=["\nObservation:"])
    parsed = parse_step(raw)

    if parsed.is_final:
        return AgentResult(final=parsed.final_answer, steps=...)

    tool = registry.get(parsed.tool)
    observation = tool.run(parsed.action_input)
    scratchpad.append((parsed, observation))

That is it. Everything in agent.py is plumbing around this.

Prompt format

Thought: <reasoning>
Action: <tool name>
Action Input: <input string>
Observation: <tool output, supplied by the loop>
... (repeat) ...
Thought: <final reasoning>
Final Answer: <concise answer>

The parser is forgiving: it tolerates case, extra whitespace, and a single extra line after Final Answer:. If the LLM fails to emit an Action at all, the loop treats the raw response as the answer (graceful degradation).

Stopping

Two ways the loop terminates:

1. Final Answer: — result.success = True
2. max_steps reached — result.success = False, the final answer falls back to the last observation.

Tuning

Symptom	First thing to try
Loop never terminates	Reduce max_steps, or use a stronger model
Tool not called	Inspect the system prompt — does the description match how the LLM thinks about the task?
Tool called with wrong args	Add a one-line example in the tool's description
Answer is verbose	Add a "answer in one sentence" instruction in your question
LLM emits two actions at once	The parser only takes the first — make the prompt mention this explicitly