Laravel AI SDK 0.11.0 introduces run-level observability for agent workflows: a single correlation ID threaded through an entire run, and lifecycle events (StartingStep, StepCompleted, StepFailed, ToolFailed, AgentFailed) with wall-clock timings. Hosted tool search now works for OpenAI and Anthropic via a ToolSearch wrapper, sparing token cost from shipping full tool catalogues on every request. Failover expands to cover connection failures, more upstream HTTP error codes (502/504/520/522/524), and Anthropic usage-cap errors. Other additions include xAI web/file search, Groq and OpenAI-compatible transcription, an assertPromptedTimes() test helper, and Gemini's default text model moving to gemini-3.7-flash. The release shipped August 19, 2026 with 36 merged PRs and twelve first-time contributors, and includes upgrade notes for two events with new required constructor arguments.
Table of contents
# Tracing a Full Agent Run# Hosted Tool Search# More Failures Now Trigger Failover# Provider Support# Testing Helpers# Other FixesQuestions this post answers
What changed with failover behavior in Laravel AI SDK 0.11?
Failover now triggers on connection failures (host down, refused connections, local Ollama not running), a wider set of upstream HTTP error codes (502, 503, 504, 520, 522, 524), and Anthropic spend-cap usage limit errors that previously threw an unmatched 400 error. A bare 500 is deliberately excluded since it can be a deterministic error that failover would mask. Track failover and error-handling changes like this on daily.dev before they break a production AI pipeline.
What breaking changes are in the Laravel AI SDK 0.11 upgrade?
Two events gained required constructor arguments: AgentFailedOver now takes a string invocationId as its last argument, and ToolInvoked takes a float time argument. Listeners are unaffected, but any code manually constructing these events needs updating. Additionally, Gemini's default text model moved to gemini-3.7-flash, so applications relying on the old default should set the model explicitly. Developers upgrading Laravel AI packages check breaking changes like these on daily.dev before running composer update.
How does the ToolSearch wrapper reduce token usage in Laravel AI agents?
ToolSearch defers a set of tools so OpenAI and Anthropic load them on demand through their own hosted search instead of shipping every tool definition on every request. Tools need no interface changes; Anthropic's search strategy is configurable between regex and bm25 via a constructor argument, and only one wrapper may be registered per request. Compare agent tool-loading strategies like this on daily.dev when optimizing token costs for AI agents.