{
  "video": {
    "id": "mJ7q8CuVmbw",
    "title": "We Trust AI Agents That Break Every Day - Shy Ruparel, Temporal",
    "duration": 1292,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "mJ7q8CuVmbw",
    "title": "We Trust AI Agents That Break Every Day - Shy Ruparel, Temporal",
    "one_liner": "Temporal's survey of 554 engineers found that 85.5% trust their agents' output while 41.1% hit agent issues daily or worse, so the aim is to make failures cost nothing, not to reach zero failures. Shy Ruparel argues that durable execution, where one event history is both the debugging record and the replay source, does that.",
    "summary": "Shy Ruparel presents Temporal's second annual State of Development survey of 554 engineers and engineering leaders who use AI agents. It shows four gaps: usage vs production, trust vs reality, prototype vs enterprise deployment, and seeing vs recovering. The top productivity blockers are tracking state and debugging, not the models. He then runs a live demo: a nine-stage payout workflow with an agentic fraud investigation, using a local Qwen model, Temporal and a human approval step. He kills the workers mid-run and shows the workflow resume from its event history without re-running LLM calls or losing the approval signal. He argues this one workflow dealt with the seven blockers named in the survey.",
    "key_points": [
      "Gap 1, usage vs production: daily agent use rose from 47.3% (2025 survey) to 80.8%, but only 49.1% run agents in production or call them core to how they ship.",
      "Gap 2, trust vs reality: 85.5% trust agent output at least somewhat, but 41.1% hit agent issues daily or more and 16.4% hit them hourly. Teams that call themselves successful are 1.7x more likely to hit issues continuously, and the number of agents a team runs is not correlated with how often issues occur.",
      "Gap 3, prototype vs enterprise: 51.3% get an AI prototype to production-ready in hours or faster. At companies with 5,000+ employees about 1 in 5 say it takes months, versus about 1 in 20 everywhere else. The speaker's read is that the extra time goes into approvals, audit trails and auditability.",
      "Gap 4, seeing vs recovering: the top productivity limiters are tracking state across services (35.7%) and isolating and debugging issues (30.9%). 22.2% lose progress to crashes and restarts, only 13% are very confident they can observe and debug AI workflows at scale, and the number one place people go for help when an agent breaks is YouTube.",
      "40% say security is what blocks them from autonomous agents, 30% complain about writing code unrelated to business logic (recovery code, state tables, retry logic, reconciliation jobs), and 8 in 10 say cost limits how they use agents.",
      "Demo: a $2,500 payout to a synthetic customer runs through a 9-stage Temporal workflow (validate, reserve funds, FX quote, an agentic fraud investigation with tool calls, a drafted brief, human approve/reject, bank submission) on a local Qwen 3.5 9B model. The workers are killed twice. The approval and bank-submission signals land while zero workers are running, new workers resume from the last successful event, and each AI activity runs exactly once.",
      "Design pattern: the deterministic workflow code owns the money and state. LLM calls, tool calls and bank API calls run as activities, and human decisions arrive as recorded signals. The agent only gives bounded advice, so if the model fails the payout continues without the brief.",
      "OpenAI's Codex and Replit's agents are cited as running on Temporal in production."
    ],
    "takeaways": [
      "Plan for agent failure rather than trying to remove it. The goal is that failures cost nothing, not zero failures.",
      "Keep the deterministic orchestration (state, money, next steps) separate from non-deterministic work: run LLM calls, tool calls and external API calls as retryable activities with recorded results, so a crash never re-spends tokens on finished steps.",
      "Bound the agent to an advisory role inside a workflow that owns the critical state, so a model outage degrades gracefully instead of breaking the process.",
      "Declare retries once as a policy instead of hand-writing retry loops, and treat bad model output as an ordinary activity failure that gets retried.",
      "Use a single event history for both debugging and recovery, so observability and recovery records don't drift apart and someone ends up reconciling state by hand at 2 a.m."
    ],
    "topics": [
      "agents",
      "reliability",
      "durable-execution",
      "state-management",
      "observability",
      "human-in-the-loop",
      "fintech",
      "surveys"
    ],
    "tools": [
      "Temporal",
      "Qwen 3.5 9B",
      "OpenAI Codex",
      "Replit",
      "KPMG",
      "Keycard"
    ],
    "quotes": [
      {
        "text": "You can't skill your way out of this.",
        "at": "06:08",
        "url": "https://www.youtube.com/watch?v=mJ7q8CuVmbw&t=368s"
      },
      {
        "text": "So the goal isn't zero failures, right? When we're using agents in production, it's just how do we make it so these failures don't cost us anything.",
        "at": "06:50",
        "url": "https://www.youtube.com/watch?v=mJ7q8CuVmbw&t=410s"
      },
      {
        "text": "Observability right now just tells you your agent failed.",
        "at": "10:15",
        "url": "https://www.youtube.com/watch?v=mJ7q8CuVmbw&t=615s"
      },
      {
        "text": "the gap between seeing uh and recovering is a person reading a trace at 2 a.m. in the morning and then fixing their state by hand.",
        "at": "10:41",
        "url": "https://www.youtube.com/watch?v=mJ7q8CuVmbw&t=641s"
      }
    ],
    "words": 4811
  },
  "summary_url": "/#mJ7q8CuVmbw",
  "transcript": {
    "html": "/transcripts/mJ7q8CuVmbw.html",
    "txt": "/transcripts/mJ7q8CuVmbw.txt",
    "vtt": "/transcripts/mJ7q8CuVmbw.vtt"
  }
}