FDE Times: Ten agent steps at 95% each leave ~60% end-to-end — cut LLM steps before polishing prompts
A 4 November FDE Times guide walks FDEs through the p^n math behind multi-step agents: at 95% per step, ten sequential LLM steps succeed only about 60% of the time (twenty steps ~36%), real accuracy can be worse as chains lengthen, and the highest-leverage fix is fewer model steps plus risk-based checks and resume-from-checkpoint flows.
FDE Times published a 4 November guide on compounding errors in agent workflows. The core point is simple multiplication: if each sequential step is correct with probability p, end-to-end success is roughly p^n. Ten steps at 95% yield about 60%; twenty steps about 36%. Mirantis and other sources cited in the article treat 95% as optimistic, and an ICLR 2026 paper notes per-step accuracy can fall as chains grow, so p^n is a ceiling rather than a forecast.
An invoice-reconciliation example shows how to redesign: move downloads, date normalisation, ERP lookups and writes into ordinary code, merge extractions into one schema call, leave two LLM steps (~90% if those two stay at 95%), add field checks against a source of truth, human gates before irreversible ERP posts, and checkpoints so failures resume instead of restarting.
For FDEs, the customer-facing rule is measure whole-flow success on real inputs from week one—not per-step demo scores—and prefer cutting model steps over adding reviewer agents that only lengthen the chain.
Source: fdetimes.net