Anti-Fragile Agentic Workflow (AFAW): A deterministic, multi-agent methodology for AI-assisted software development

Coding agents based on large language models can produce code faster than a person can review it. Running several of them on one repository adds failure modes that a single-agent workflow does not have: agents working on the wrong branch, collisions on shared context files, claims of results that were never measured, and tests that pass whatever the code does. This paper describes the Anti-Fragile Agentic Workflow (AFAW), a repository-level methodology and boilerplate for AI-assisted development with several agents in parallel. AFAW rests on one division of labour: the AI decides, the engine measures. Agents generate code; deterministic tools (tests, type checkers, mutation testing, continuous integration) decide whether it is acceptable; a human approves every merge. The method combines per-task branch isolation, task-scoped state files merged by continuous integration, strict test-driven development, validation in parallel cloud jobs, mutation testing restricted to modified files, and enforcement layers that do not depend on the prompt. The reference implementation targets a Python backend; the stack-specific rules are meant to be adapted to other stacks. The rules were derived from failures the author observed in practice. This paper states the method, maps each failure mode to a mechanism and to its residual risk, lists the limitations, and proposes an evaluation protocol. It reports no performance measurements: whether the method improves outcomes remains to be measured.