A Minimal ReAct Loop: Thought -> Action -> Observation

The problem, from first principles

Every "agent" that does more than answer one question in one shot is running some version of the same tiny loop: the model reasons about what to do next (a thought), decides on a concrete action to take, and gets back an observation — the real-world result of that action — which feeds into its next thought. Repeat until the model decides it has enough to answer. This loop, ReAct (Reason + Act), is the base pattern underneath almost every agent framework; everything more sophisticated (retries, routing, multi-agent handoffs) is a variation layered on top of this same three-part cycle.

From theory to code

You're given the model's full scripted sequence of (thought, action, action_input) turns — as if already generated, no real model call needed here — and a lookup table of what each action call actually returns. Implement run_react_loop(steps, observations, max_steps). Process steps in order, up to max_steps. For each step, look up its observation from observations (using "NO_OBSERVATION_FOUND" if that exact (action, action_input) was never scripted), and append (thought, action, action_input, observation) to the trace. The moment action == "finish", stop immediately — record that step with an empty observation and don't consult observations at all for it.

Return the trace built so far.

Constraints

• 1 to 50 steps in the log; max_steps >= 1.
• action is never the literal string "finish" except to signal the loop should stop.

Hints