We Trust AI Agents That Break Every Day - Shy Ruparel, Temporal

하이라이트

• Corporate productivity gains from agentic AI reached 3 trillion dollars in 2025, with financial services leading at 94% adoption.

• Daily agent usage doubled from 47.3% to 80.8% over the course of a year, while only 49.1% of teams operate agents in production.

• Over 85% of engineers trust agent output at least somewhat, yet 41.1% encounter system issues daily and 16.4% experience them hourly.

• Tracking state and debugging rank as the top two blockers for engineering team productivity, bypassing model limitations entirely.

• Durable execution allows code to run to completion even when processes crash, preserving state and event history without requiring custom recovery code.

타임라인

1. 00:34 - 03:16

   Corporate AI Adoption and Industry Gaps

   Financial operations run on long-running workflows spanning multiple systems and days, involving real money that leaves no room for error. A survey of engineers with substantial industry experience highlights that despite high adoption rates, a significant operational gap separates daily use from reliable production deployment.

   • Corporate productivity gains from AI agents totaled 3 trillion dollars in 2025 according to a KPMG survey.
   • Financial services lead adoption with 94% of fintech firms treating AI as core to operations.
   • A survey of 554 experienced engineering professionals revealed four major gaps in agent deployment.
2. 03:17 - 07:18

   Usage Versus Production and Trust Versus Reality

   A stark contrast exists between how widely agents are used and how often they successfully reach production. High confidence in agent outputs clashes directly with frequent system failures occurring on a daily or hourly basis, proving that technical skill alone does not eliminate system instability.

   • Daily agent usage jumped from 47.3% to 80.8%, but only 49.1% of teams run agents in production.
   • Eighty-five percent of surveyed engineers trust agent output at least somewhat.
   • Fourteen percent of teams encounter agent issues daily, and 16.4% experience system failures every hour.
3. 07:18 - 13:10

   Deployment Speeds and Reliability Obstacles

   Large regulated enterprises require strict audit trails, approvals, and guaranteed transaction handling, which extends deployment timelines to months. State management and debugging emerge as primary operational bottlenecks, forcing engineers to write extensive auxiliary code unrelated to core business logic.

   • Half of all teams transition an AI prototype to production in hours, whereas large enterprises with over 5,000 employees often require months.
   • Tracking state across services limits productivity for 35.7% of teams, while 30.9% struggle with debugging.
   • Twenty-two percent of survey respondents lose progress and must repeat work due to application crashes.
4. 13:10 - 21:24

   Durable Execution and Live Demonstration

   A live payout workflow demonstration shows how killing background workers mid-process fails to disrupt ongoing transactions. Signals such as human approvals are safely captured and processed upon worker recovery, eliminating the need to waste expensive LLM tokens or rebuild application state from scratch.

   • Durable execution ensures workflows run to completion even when individual workers crash.
   • Temporal records every state change, tool call, and human approval signal in an immutable event history.
   • Production environments like OpenAI's codex and Replit utilize durable execution to prevent catastrophic workflow failures.