Answer
What makes a long-running AI job fail?
Usually something mundane: a limit, an expiry, a machine going to sleep, or a cost cap nobody set.
Operational things more often than reasoning things: a rate limit, a credential expiring mid-run, the machine sleeping, a network interruption, or a cost that nobody capped. All are preventable and none is about the model.
Discussion of long-run failure concentrates on drift and context, which are real and are not what usually stops a job. The failures that actually occur are the ones any long-running process has: something hit a limit, something expired, something went to sleep, something disconnected. These are unglamorous, they are entirely preventable, and they are worth checking first because the diagnosis is fast and the fix is permanent.
Rate limits are the most common. A run making many calls will meet a per-minute or per-hour ceiling, and what happens next depends on whether anything handles it. With a retry and a backoff the run pauses and continues; without one it fails, often partway through a stage, leaving state that may or may not be consistent. Knowing which of those your setup does is a five-minute check.
Credential expiry is the second and produces the most confusing failures, because the run works for two hours and then does not. A token with a lifetime shorter than the job will expire mid-run, and the resulting errors look like the service being unavailable rather than like an authentication problem. Checking the lifetime against the expected duration before starting is the whole of the prevention.
The machine is the third and applies to anything running on a laptop. Sleep settings, screen lock, power management, a scheduled update, someone closing the lid. A long run needs a host that stays awake, and for anything genuinely long that argues for somewhere other than the machine a person is carrying around.
Cost is the fourth and is the one that fails silently in the other direction: the job does not stop, it continues, and the failure is discovered on the bill. A run that has entered a retry loop or an unproductive investigation will keep spending, and only a cap stops it. This is the reason to set one even when the expected cost is small, since the expected cost is not what caps are for.
Disk and output size deserve a mention because they are easy to miss. A run producing logs, intermediate files or downloaded material can fill the space available, and the resulting failures are strange — a write fails, something is truncated, the state file is written partially. Checking what a run accumulates and where is a small task that avoids a puzzling afternoon.
The agent did not get confused; the laptop closed.
Siddharth Sharma, Context Theory
Related questions
How do you tell an operational failure from a reasoning one?
The log, and specifically whether the last entry is an error from a tool or a decision by the agent. An operational failure stops abruptly at a call; a reasoning failure produces a completion report that is wrong. If the run ended without a report, look at the operational causes first, because they are faster to rule out.
What should a long run do when it hits an interruption?
Write its state and stop cleanly, which requires the state to be written continuously rather than at the end. A run that saves progress only on completion loses everything to any interruption, and interruptions over several hours are close to certain rather than exceptional.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
| Firms that never responded to a web enquiry at all | 23% | Category-wide |
2026 speed-to-lead benchmark · verified
Oldroyd, McElheran & Elkington, "The Short Life of Online Sales Leads", Harvard Business Review (March 2011) · 1.25M inbound leads across 2,241 US firms · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Software | Long agent runs fail predominantly for the reasons any long-running process fails — rate limits, credential expiry, host sleep, network interruption and unbounded cost — rather than for reasons specific to language models. | Examining the final log entries of failed long runs and classifying the cause of each. |
| Workflow | A credential whose lifetime is shorter than the job expires mid-run and produces errors resembling service unavailability rather than authentication failure, which makes comparing token lifetime against expected duration a specific pre-run check. | Checking the expiry time of the credentials in use against the expected run duration. |
| Buying behaviour | Cost failures do not stop the run, so a job in a retry loop or an unproductive investigation continues spending until a cap intervenes, which is why caps exist for the unexpected case rather than the expected one. | Comparing the cost of runs that completed normally against those that entered a retry loop. |
| Response | A run that writes state only on completion loses all progress to any interruption, and over several hours interruptions are close to certain, which makes continuous state writing a requirement rather than a refinement. | Interrupting a long run partway and checking what state survives. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one