ReAct: Reason, Act, Observe

Think, then do, then look

ReAct (Reasoning + Acting) interleaves a short thought, an action (a tool call) and an observation (the tool result), then repeats. Thinking about the observation before the next action helps the model correct course. Modern tool-calling APIs give you this loop structurally; the model can also reason before calling a tool.

A ReAct transcript

Older text-only agents asked the model to print these lines and parsed them. The parsing line below ran as shown: it pulls the tool name and JSON arguments out of the text.

text = ("Thought: I need the balance.\n"
        "Action: get_balance\n"
        "Action Input: {\"account_id\": 7}")
fields = dict(l.split(": ", 1) for l in text.splitlines())
print(fields["Action"], json.loads(fields["Action Input"]))

Output:

get_balance {'account_id': 7}

Prefer structured tool calls

Parsing free text breaks when the model changes wording. Native tool calling returns a validated name and JSON arguments, which is far more reliable than regular expressions on prose.

Quick check: What happens after an action in ReAct?

• The model sees an observation and reasons about the next step
• The run ends
• The tools are deleted
• The user must restart