Reflection

The Reflection block uses an LLM to critique or synthesize outputs of another block (the “generator” block).

There are two modes:

• Critique & Refine - The generator block runs, and a judge LLM evaluates the result against the judge instructions. With Improve with feedback, the critique is passed back into the next iteration to revise the answer. With Simple retry only, the generator block is run again without that feedback.

• Synthesize - The generator block runs several times, then a synthesizer LLM merges the candidate outputs into one final answer. These runs can be parallelized.

In critique mode, the block can either stop with an error when no candidate meets the expectations, or continue through a failure transition with the last candidate output.

Limitation

Tool calls that require human validation are not supported inside Reflection block.