Cloudflare's Agents SDK moves to Pi 1.5, and it is a one-way upgrade
Agents SDK 0.33.0 batches storage writes instead of writing on every update, cutting SQLite transactions up to 27x, but a conversation written on Pi 1.5 cannot be read back by Pi 1.4.
Cloudflare shipped Agents SDK 0.33.0 on 3 November, upgrading its Pi harness from 1.4 to 1.5. The change is about where the time and memory go once an agent is actually running.
Pi used to write streamed answers and tool output to storage on every update. 1.5 batches those writes, within a second of each change, which Cloudflare says cuts storage transactions by about 27x for long answers and 5.7x for turns with parallel tools, with roughly a quarter less CPU. A benchmark with 2,048 SQLite sessions went from 88 to 459 agent runs per second. Large streaming responses also switch from copying text to appending it, which Cloudflare says took a 500KB tool call from around 100MB of heap down to far less.
The catch is that this is a one-way door: once a conversation has run on Pi 1.5, Pi 1.4 cannot read it back, so a deployment cannot roll back past this release. session.events() is replaced by session.watch(), retry() and the parked status are gone, and there is a new session.continue() for resuming a run after a model error without appending a message.
Why it matters: if you’re running agents on Cloudflare’s Pi harness, this is a real throughput and cost win, but it needs planning rather than a routine dependency bump. Read the breaking-changes list before you deploy, since there is no going back once live traffic has written state in the new format.