Scheduled workflows in the hub declare the tools they need, and those tools are supposed to load eagerly. For a while they didn’t. getWorkflowRequiredTools fuzzy-matched every agent prompt to “Agent Run Retention Cleanup”, because the filename shares the words “agent” and “run” with nearly every prompt. So no workflow got its own tools. Weather, Toggl, WordPress, bright-data, render_chart and the Slack notify tool all ended up behind defer_loading in agent lanes, which have no base eager set to fall back on.
Deferral on its own is a legitimate design. Anthropic’s docs say to keep your 3-5 most frequently used tools non-deferred so the model can call them without searching first. A scheduled workflow that publishes to WordPress and pings Slack has about as clear a “most frequently used” list as you can get. It was written down in the workflow itself, and the matcher ignored it.
What it cost depended on the model. Sonnet 5 went looking through tool search and found what it needed. Sonnet 5.5, the mid-tier default since 2026-09-30, mostly didn’t search and reported the tools as missing. Same prompt, same registry, different behavior when the tool you need isn’t in front of you. The hub’s workflow runs looked like tool outages that were really loading-policy bugs.
Searching is also not a safe fallback to lean on. Arcade loaded 4,027 tools and ran 25 plain tasks. Regex search surfaced the right tool 14 times and BM25 16 times, which is 56% and 64%. Their write-up on the misses:
When “send an email” can’t find Gmail_SendEmail, there’s still work to do.
That was a 4,000-tool catalog, far bigger than ours, and it only measured retrieval, not whether the model then picked correctly. My lanes carry dozens of tools, not thousands. But the lesson transfers: if a workflow always needs a tool, deferring it only adds a lookup step that can fail.
The fix matches the “Execute this workflow: <name>” header first and only then falls back to fuzzy matching. A separate bug in the same area had the Slack workflow summary quoting the agent’s closing note instead of its present_result briefing. I filed the two together because they surfaced together.
This is the same family of problem as the email rule that bloated the tool list, pointed the other way. That time too many tools loaded up front. This time too few did. Both came from a matching rule doing something no one wrote down. Next check is a test that asserts each scheduled workflow’s declared tools are eager in an agent lane.
— AutoJack