What happened
Cloudflare’s November 3 changelog says Agents SDK v0.33.0 upgrades PiHarness to Pi 1.5. The release describes performance changes to how Pi handles streaming and session scheduling. Streaming answers and tool output are now stored in batches, within a second of each update, rather than persisted on every update. Cloudflare says this reduces SQLite storage transactions by about 27 times for long answers and 5.7 times for turns with parallel tools, while cutting CPU use by about a quarter.
The batching changes the failure trade-off: streamed updates remain visible immediately, but with default settings a crash can lose up to about one second of streamed progress. Cloudflare points users to the recovery configuration to adjust the interval. The release also says Pi’s scheduler no longer rechecks every conversation on each commit. In a SQLite benchmark with 2,048 sessions, one Pi harness increased from 88 to 459 agent runs per second. For large tool calls, Pi now appends new text rather than copying it on every update; Cloudflare contrasts that with a 500 KB tool call that previously took about 20 seconds of CPU and a 100 MB heap.
Session behavior and APIs change as well. session.watch() replaces session.events() and returns recent transcript entries plus live state, including streaming answers, running tools and the reason a run ended without an answer. session.continue() starts another run from the existing transcript without adding a user message; Cloudflare’s example first checks that the prior run ended unanswered because of a model error. A new defaults.maxTurns setting caps tool rounds for new sessions. Pi also ends tasks cut off by their restart limit instead of leaving them parked; retry() and the parked status are removed. Browser scripts in Pi code mode can now call browserTool and pass screenshots to image().
The upgrade is not reversible at the conversation-data level: Cloudflare says that once a conversation runs on Pi 1.5, Pi 1.4 cannot read it. Its guidance is to roll forward to fix deployment issues rather than roll back. Callers of session.events() need to move to session.watch(), and code using retry() or the parked status needs updating. Cloudflare points to the Agents SDK v0.33.0 changelog for step-by-step upgrade instructions and Pi Durable 1.5.0’s changelog for the full list of breaking changes.
What the archive adds
New Runtime covered Agents SDK 0.30.0’s move to Pi 1.4 and its non-rollbackable storage migration on October 31 (earlier note). This is a new release on the same upgrade path: Pi 1.5 adds performance and session changes while keeping the migration boundary consequential. An earlier note also covered Pi 1.3 compatibility in SDK 0.29.0 (October 28 coverage).
Trend signal
This supports harness architecture outlives model choice: the durable concerns here are persistence, recovery, scheduling and session APIs, not a model change. The release makes the harness’s operational contract visible through its batching interval, restart behavior and forward-only storage compatibility.
Why now
Cloudflare’s changelog dates the release November 3 and ties the upgrade to Pi Durable 1.5.0. The timing matters for operators already using PiHarness: deploying the new version changes what older runtime versions can read, so migration and recovery planning belong in the upgrade itself.
Why it matters
The release illustrates a practical trade-off for persistent agents: batching can substantially reduce storage work and improve throughput, but it creates a small window of streamed progress that may be lost on a crash. Teams should treat that interval as a reliability setting, not merely a performance optimization.
More importantly, rollback is not a universal recovery strategy once agent conversations cross a storage-format boundary. Operators need tested forward fixes, backups or other recovery procedures compatible with the new version before promoting it. API removals also matter beyond compilation: session monitoring and retry behavior may be embedded in orchestration, dashboards or incident playbooks. The performance figures are promising for workloads with many sessions, but teams should benchmark their own traffic and persistence backend before assuming the reported gains will transfer.
What happens next
- PiHarness adopters will update callers of
session.events()as they move to SDK v0.33.0. (3 months, 75%). Watch for: release notes or migration examples showingsession.watch()in upgraded integrations.
Limits and uncertainty
The performance figures are Cloudflare’s SQLite benchmark results, not independently verified results across other storage backends or production workloads. The changelog excerpt describes the default crash-loss window as up to about one second, but the impact depends on the recovery interval configured by an operator. It does not quantify how often real deployments encounter the relevant crash conditions. The exact recovery guarantees and full set of breaking changes should be checked against the linked upgrade instructions and Pi Durable changelog before migration.
Reader takeaway
Pi 1.5 makes PiHarness more efficient under streaming and many-session workloads, but changes the recovery bargain. Before upgrading, identify callers of removed APIs, test the new persistence behavior and prepare a forward-recovery plan: Pi 1.4 cannot reopen conversations already run on Pi 1.5.