Source: github.com/cloudflare/agents/releases/tag/agents%400.34.0
The Pi 1.6 support covers the agents/harness/pi, agents/models/pi-ai and Pi tools. A tool can now wait for a signal or a deadline while the agent sleeps, signals are delivered with harness.signal(), and submit() and continue() accept notify so a parked tool wakes when another session's input settles. Queued input can be kept across an abort with onAbort: "keep".
MCP clients now throw MCPAuthorizationRequiredError with a link when a request needs the user to sign in again, and a failed server stays failed until reconnectMcpServer(id) is called. Failed requests no longer expose error details: HTTP requests get a generic 500 and WebSocket upgrades close with 1011, and onFetchError() can choose the response.
Other changes include FLUX.2 image models in ai.image() and ai.images(), claudeCode() defaulting to claude-sonnet-5-5, and cancelled schedules, queue items and Tasks no longer rerunning.
Why it matters if you build agents
Pi 1.6 is a one-way upgrade: once a conversation runs on Pi 1.6, Pi 1.5 cannot run its tool calls or model requests, so a bad deploy has to be fixed by rolling forward.
Code that catches UnauthorizedError keeps working because MCPAuthorizationRequiredError extends it, but code that calls callTool(), readResource() or getPrompt() should send users to error.authUrl. Apps that relied on error details in responses should implement onFetchError.
Get stories like this every weekday
One short email with the day's agent tooling news.