Shy Ruparel, a developer advocate at Temporal, presents a survey and a payment-workflow demo in this 22-minute AI Engineer talk.

Four gaps between adoption and dependable systems

  • Temporal surveyed 554 engineers and engineering leaders already using AI agents; the results describe that selected group.
  • Daily use reached 80.8%, while 49.1% reported production use or agents central to shipping.
  • Some 85.5% trusted agent output at least somewhat, even as 41.1% reported issues daily or more often.
  • Large-company deployment delays raised questions about approvals and auditability; Ruparel labels that explanation as his interpretation.
  • Tracking state across services led reported productivity blockers at 35.7%, followed by debugging at 30.9%.

Keep business state outside the worker

  • His synthetic $2,500 payout reserves funds, runs an AI fraud investigation, and waits for human approval.
  • A worker crash can otherwise leave funds reserved in one system and the approval brief lost in another.
  • Temporal records completed work and incoming events in a persistent history that survives workers stopping.
  • Workflow code decides the next step deterministically; model calls, tool calls, and bank calls run as recorded activities.

Recovery has a precise boundary

  • Completed activity results can be reused without another model call or token charge.
  • An activity interrupted before its result is recorded can retry. External side effects still need duplicate-safe handling.
  • Human approval arrives as a persisted signal, including while no worker is running.
  • The agent advises on the transaction; workflow logic owns the money and state, with a human making the approval decision.

What the demonstration actually showed

  • The live restart stalled, and Ruparel switched to a recording showing the intended crash-and-resume sequence.
  • He reads its history as showing each completed AI activity ran once, with approval and bank-completion signals surviving worker stoppages.
  • The same history supports debugging and recovery; durable execution alone does not make a model’s judgment correct.

“The agent itself gave us advice. The workflow itself owned the money and the state.” — Shy Ruparel, 19:19

Caveat: the survey is vendor-sponsored and the payout uses a synthetic customer. The successful recovery sequence is shown in a recording after the live restart stalls.