Temporal CEO on AI Agents & The Future of Software | Deep Dives with a16z

影片總結

Temporal powers OpenAI Codex and Snap at scale, bringing durable state, recovery, and RPC to long-running AI agents.

Temporal CEO Samar Abbas explains how durable execution records application state, resurrects failed workflows on another host, and guarantees completion despite distributed-system chaos. The platform grew from Cadence at Uber, where it handled tipping retries lasting up to 3 days, loyalty workflows, and recovery from a bug that reset customer points. Abbas says Temporal now powers OpenAI Codex and Snap’s story processing, with its cloud handling bursts of 150,000 actions per second. As agents shift from interactive prompts to asynchronous, long-running work, they need recoverability, sandboxes, observability, context engineering, and durable RPC. Temporal’s execution histories provide auditability and analytics, while Project Nexus aims to standardize asynchronous tool calls between specialized agents. Abbas argues SaaS is not dead: value will accrue to APIs, reliable infrastructure, and breakthrough applications such as Harvey and healthcare tools.

章節

1. 0:00Introduction: Durable Execution Guarantees Exactly-Once Order Processing
2. 4:03Temporal’s Origin Story: Cadence Solved Uber’s Microservice and Loyalty-State Failures
3. 11:14Why Agents Raise the Stakes: AI Reliability Requires Multi-Region Continuity
4. 16:00Specialized Agents Need Durable RPC: Background Agents Move Beyond the MS-DOS Era
5. 20:01Specialized Agents Need Durable RPC: Codex Orchestrates Long-Running Coding Work
6. 25:20Deep Research Agents: Parallel, Human-in-the-Loop Workflows Need Recovery
7. 30:58Execution Histories as a Superpower: 100x Scale and Auditable Agent Traces
8. 35:03Execution Histories as a Superpower: APIs Preserve Value as Software Expands
9. 39:04Minimal Viable Long-Running Agent Architecture: Sandboxes, Evaluation, and Systems of Record
10. 45:07Context Engineering at Scale: Real-Time Retrieval and Durable RPC for Agent Swarms
11. 52:40Where Value Accrues: Jensen Huang’s Five-Layer Cake and Breakout AI Applications
12. 56:01Where Value Accrues: Temporal’s 2021 Lessons, Margins, and R&D Expansion