ReAct: Reason + Act Loops That Actually Work
Why ReAct became the default
ReAct (Reason + Act) is popular because it forces a healthy rhythm: the model thinks, it takes an action (often a tool call), it observes the result, then it thinks again. That observation is the antidote to hallucination.
The ReAct loop in plain language
Instead of “think once → answer”, ReAct encourages:
Thought → Action(tool) → Observation → Thought → Action → ... → Final Answer
In production you usually hide the raw “thought” from users, but you keep the loop internally.
Designing actions that are safe
Actions should be typed and constrained. For example, don’t allow “run arbitrary shell command”. Instead allow “searchDocs(query)”, “getOrderStatus(orderId)”, “createTicket(summary, priority)”.
Think of tools as an API contract your agent must respect.
A practical prompt skeleton
A reliable pattern is:
- System: role + safety rules + tool usage policy
- Developer: response format + stop conditions
- User: goal + constraints
- Memory: relevant context only
Common failure: endless tool calls
Agents can fall into “tool addiction”: they keep calling tools because it feels productive. Fix with:
- Max tool calls per run
- Budget tracking
- Stop criteria (“If confidence ≥ X, answer now”)
- Tool choice policy (only call tools when needed)
Production tip: keep observations small
Tool outputs can be huge (HTML pages, logs). Summarize and store the summary. Feed the LLM only what it needs to decide next actions.
