Agent 基础设施官方公告自动监测

Agents - Agents SDK now supports Pi 1.5 with fewer storage writes and lower memory use

Cloudflare Agents SDK v0.33.0 upgrades PiHarness to Pi 1.5, significantly reducing storage writes and memory use during streaming while introducing session management updates.

原始内容为英文;当前页面提供中文导航与来源说明,具体事实请以原文为准。

人类阅读

为什么值得关注

Cloudflare Agents SDK v0.33.0 upgrades PiHarness to Pi 1.5, introducing substantial performance and architecture changes. Persisted storage transactions are reduced through batching streamed answers and tool outputs within a second of updates rather than per-update, lowering SQLite transactions and reducing CPU usage by roughly 25%. In benchmark testing with 2,048 sessions on SQLite, agent throughput rose from 88 to 459 runs per second due to scheduler optimizations that avoid rechecking all conversations on commit. Memory overhead is mitigated by appending streamed tool call text instead of copying per update. Crash recovery defaults trade persistence frequency for efficiency, retaining an up to ~1-second data loss window unless reconfigured. New APIs include session.watch() replacing session.events() , session.continue() for resuming runs from existing transcripts without new messages, defaults.maxTurns for capping run tool rounds, browserTool support in code mode with abort handling, and the termination of repeatedly failing tasks, removing retry() and the parked status.

Agent 解析

可执行摘要

Agents SDK v0.33.0 updates PiHarness to Pi 1.5, implementing batched streaming persistence (approx. 1-second batching), string appending to cut memory usage, and scheduler optimizations. Breaking and API changes include replacing session.events() with session.watch() , adding session.continue() , adding defaults.maxTurns , and removing retry() and the parked status.

Agent 实用度
80/100
可信度
90%
机器格式
JSON + Markdown
下一步

开发者应核对什么

  • Upgrade Agents SDK to v0.33.0 if using Cloudflare Agents with PiHarness.
  • Replace occurrences of deprecated session.events() with session.watch() .
  • Remove references to retry() and the parked status, as tasks exceeding restart limits now terminate completely.
  • Review crash-recovery settings if the default 1-second batched persistence interval is unacceptable for your use case.
分类

标签与路由

cloudflareagentschangelog
相关信号

继续阅读