Not literally. But every box you draw becomes a number on a spreadsheet within a quarter, and with AI agents the conversion is harsher than anything I’ve seen in more than two decades of SaaS/Software.
Here’s the translation list:
No durable execution → rework. Crash at step eight of ten, re-run from the top, pay for seven steps twice. Multiply by your deploy frequency. That’s a real line in the token bill: the silent tax.
Side effects with no completion record → reconciliation. If the runtime can't tell whether the refund step finished before the crash, it either skips it or fires it twice. Either way, someone in finance cleans it up by hand, and their time costs more than your tokens.
No execution record → audit overhead. Every compliance question becomes an engineering archaeology project billed by the hour.
Framework lock-in → switching cost. The framework you commit everything to today has a real chance of being the one you migrate off next year.
The metric that fixes the conversation: cost per completed task, not cost per token. Cheap tokens on an unreliable runtime lose to expensive tokens on a durable one, because durable workflows only pay for work once while recovery replays the journal instead of re-running the model.
Quick illustration: a 10-step agent, roughly 1k tokens per step. It crashes at step eight, crashes again on the retry, then finishes on the third attempt. Restarting from the top re-pays steps one through seven each time, so that's 14k tokens of rework on a 10k-token task. Your cost per completed task is 2.4x what the happy path suggests. With durable resume, steps one through seven are paid for once, and the task costs roughly what it should.
Now scale it to a fleet. If 15–30% of workflows fail, as is common in an early deployment, restart-from-zero can add up to that much again to your model bill, and it never shows up as its own line item.
When your architecture makes that metric flat and predictable, budget meetings become short. When it doesn’t, they become frequent.
The full math, worked through with real failure rates: How Durable Execution Cuts Rework and Token Spend and The Economics of Running AI Agents at Scale.
Top comments (0)