๐Founder, AFP / SafeLoop ยท AI workflow reliability and governance
AFP: from prompt governance to AI workflow reliability
2026 positioning
Antifragile Prompting Framework (AFP) began with prompt governance: making complex LLM tasks more controllable, reproducible, and explainable. As AI systems move from single-turn answers toward agents, tool use, and multi-step workflows, AFP now focuses on a broader question: how can important AI outputs be verified, traced, corrected, and tested again?
AFP does not place reliability in one better prompt. It structures assumptions, evidence, counter-evidence, validation, failure handling, and revision into a workflow that can be inspected. The aim is not to claim that AI is always correct, but to make errors easier to detect, contain, and review.
SafeLoop is AFP’s core feedback cycle. It treats an output as work that remains subject to verification rather than as a final answer by default. When models, tools, or external evidence change, the same workflow should be runnable again and its results compared.
This site is being developed as the public reference for AFP specifications, evaluation methods, failure cases, and reproducible tests so that claims about reliability can be judged by evidence rather than by positioning alone.
AFP structures six things that increasingly matter once AI systems can search, call tools, use memory, and act across multiple steps:
Assumptions โ Evidence โ Counter-evidence โ Decision โ Validation โ Revision
The framework is designed around inspectability: important claims should have a traceable basis, failure should be detectable, and revised outputs should be testable again.
The next phase of AFP is evidence-first. The public reference is organized around three surfaces:
- Specification โ what AFP currently requires.
- Evaluations โ how AFP should be compared against simpler baselines.
- Failure Cases โ concrete ways AI workflows fail and how the framework responds.
AFP’s public documents, whitepaper, and subsequent research outputs.