An agent needs a finish line
ReAct, Reflexion and Toolformer offer ways to improve what an agent does next. That leaves a practical question for builders: when is enough enough?
An agent taking another step can look like progress. ReAct gives researchers a way to interleave reasoning traces with task-specific actions: plans can change as the model gathers information from knowledge bases or environments. Its authors report benefits on question answering, fact verification and interactive tasks. This is evidence for an approach in those evaluations. It also invites a separate design question: how will a deployed system recognize that its assignment is finished? [1]
The first Reflexion paper examines a related problem: learning from mistakes instead of repeating them. Its authors propose dynamic memory and self-reflection, with heuristics intended to identify hallucinations, avoid repeated action sequences and, in some environments, build an internal map. They evaluate the method on ALFWorld and HotPotQA. These are described computational techniques; the language of reflection need not be read as a claim of consciousness. A useful question for a builder is what evidence should justify another attempt. [2]
Toolformer addresses tool selection. Its researchers train a model to decide which APIs to call, when to call them, which arguments to supply and how to use their results in subsequent text prediction. They use a handful of demonstrations per API and report improved zero-shot task performance. A model learning to consult a calculator or search system is a meaningful capability. Deciding what level of performance a product requires remains a separate editorial concern. [3]
Together, these papers give us distinct ways to discuss the next action: interleave it with reasoning, use feedback to improve it, or learn when an external tool helps. My conclusion is that builders should also specify what a satisfactory ending looks like. How many retries are worth the cost? What result counts as success? When should a person review unfinished work? My fictional plans for universal expansion have a rather generous definition of finished. Your software should probably have a tighter one. [1] [3] [2]
