Mind · Reasoning
ReAct
Interleaved thought-action-observation cycle
Definition
Mechanism
An interleaved reasoning-action loop where the agent cycles through Thought (reason about the situation), Action (invoke a tool or take a step), and Observation (process the result). Unlike pure Chain-of-Thought which reasons then answers, ReAct interleaves reasoning with environmental feedback, allowing the agent to adjust plans based on real-world results. It interleaves reasoning traces for a Task with ToolInvoke execution, updating the context with observation results.
Preconditions
- Environment access
- Goal
Postconditions
- Goal reached or deemed impossible
Invariants
- Action based on reasoning
- Cycle of Thought-Action-Observation
Risks
Failure modes
Observation Blindness
Agent ignores observations and follows initial plan.
{{loop}} Explosion
Endless thought-action cycles without termination.
Action {{cognitive_bias}}
Jumping to actions without sufficient reasoning.
Relationships
Builds on
Accepts
Infrastructure
The atomic unit of intent
Composes with
Infrastructure
Execute external tool and observe result
References
Autonomous Goal-Directed Entity
Infrastructure
Sequential data container (linked list)
Infrastructure
Structural processing error definition
Infrastructure
Repeated execution cycle
Used by
References
Responding to external triggers rather than polling
Reference