ReAct

ReAct (Reason + Act) is an agent pattern where the model alternates between thought and tool-use steps — reasoning about what to do, calling a tool, observing the result, reasoning again.

ReAct was the first widely-adopted agent loop. The model emits a 'Thought' (reasoning), then an 'Action' (tool call), receives an 'Observation' (tool result), and loops until done. Most modern agent frameworks are direct descendants of the ReAct paper.

Example

Thought: 'I need to find the user's last order.' Action: `get_orders(user_id=42, limit=1)`. Observation: order #1234, $89.50. Thought: 'Now I can answer their refund question.'

How OpenLabor uses it

OpenLabor's agent runtime extends the ReAct loop with planning, retries, and approval gates.

Is ReAct still state of the art?

It's the foundation. Modern agents are ReAct-plus: planning, multi-agent coordination, structured output.

Related: agentic-workflow, tool-use, function-calling.

AI Labor Glossary