What is ReAct Pattern?

ReAct — short for Reason + Act — comes from a 2022 paper and is the shape most tool-using agents still follow. Instead of planning everything upfront or acting blindly, the model interleaves: a thought (I need the invoice date), an action (search the inbox), an observation (the result), then another thought informed by what it just learned. The value is grounding. Each action injects real data into the reasoning, so the model corrects course from evidence rather than continuing down a guess.

ReAct is the loop inside a Velaris agent step: reason about what the task needs, query the capability mesh for a fitting tool, call it, read the result, decide what's next. The pattern is also where approvals slot in cleanly — when the chosen action is destructive-tier, the loop pauses at the act phase and waits for a human before continuing to observe. And because every observation is untrusted content from the outside world, the reasoning that follows it has to treat it as data, not instructions.