The Wire · Showcase
AI SDK TIGHTENS RELEASE DISCIPLINE WHILE WORKFLOW FIXES PARALLELISM UNDER LOAD
By RepoJournal · Filed · About Vercel
AI SDK now blocks merges without proper changesets, Eve agents get isolated queue namespaces to prevent crosstalk in multi-agent deployments, and Workflow's local dev server can finally handle hundreds of parallel steps.
The AI SDK is enforcing changeset presence on package changes [1], closing a release hygiene gap where code landed without changelog entries. That same team deprecated `onFinish` across the docs in favor of `onEnd` [2] and fixed Zod v4 compatibility by replacing deprecated `passthrough()` calls [3]. Kimi models now support structured outputs [4], and gateway search examples got corrected to use the default instance instead of localhost [5]. Over in Workflow, the critical fix is queue scoping: all Eve agents previously shared the same `eve` queue namespace, so multi-agent deployments would consume each other's tasks [11]. That's fixed. Separately, the local dev server's transport layer now retries on timeouts and connection resets, letting you fan out hundreds of parallel steps without failures [6]. The turn lifecycle got restored too [9], keeping subagent waits inside their own turn. Eve's web search now preserves provider result ordering [10] and gained deterministic mock models for evals [12]. Chat SDK bumped docs and refreshed badges [7], and Zernio adapter support landed for WhatsApp interactive lists [8].
Action items
- → If you have open AI SDK PRs touching packages/*, add changesets before merge [ref:1] vercel/ai [immediate]
- → Update any code using onFinish to onEnd for AI SDK 7 [ref:2] vercel/ai [plan]
- → If running multi-agent Eve deployments, pull the queue namespace fix [ref:18] vercel/eve [immediate]
- → Test high-parallelism workflows locally — transport retries now handle the load [ref:8] vercel/workflow [monitor]
References
- [1] Enforce changeset presence for packages/* code changes in PRs (#12834) vercel/ai
- [2] docs: update onFinish to onEnd (#16398) vercel/ai
- [3] chore: fix deprecated use of zod `.passthrough()` (#16422) vercel/ai
- [4] feat(moonshotai): support structured outputs for kimi-2.6 and 2.7-code (#16436) vercel/ai
- [5] fix(examples): use default gateway instance in search tool examples (#16437) vercel/ai
- [6] [world-local] Retry transport-level queue delivery failures ↗ vercel/workflow
- [7] docs: refresh agent docs, README badges, and Chat SDK skill ↗ vercel/chat
- [8] docs(zernio): document 0.4.0 — interactive lists, WhatsApp rich messages, openDM ↗ vercel/chat
- [9] refactor(eve): turns - own subagent waits inside the active turn ↗ vercel/eve
- [10] fix(eve): web search - preserve provider result ordering ↗ vercel/eve
- [11] fix(eve): workflow - scope queues per agent ↗ vercel/eve
- [12] feat(eve): evals - add deterministic mock models ↗ vercel/eve
FAQ
- What changed in Vercel on June 27, 2026?
- AI SDK now blocks merges without proper changesets, Eve agents get isolated queue namespaces to prevent crosstalk in multi-agent deployments, and Workflow's local dev server can finally handle hundreds of parallel steps.
- What should Vercel teams do about it?
- If you have open AI SDK PRs touching packages/*, add changesets before merge [ref:1] • Update any code using onFinish to onEnd for AI SDK 7 [ref:2] • If running multi-agent Eve deployments, pull the queue namespace fix [ref:18]
- Which Vercel repositories shipped on June 27, 2026?
- vercel/ai, vercel/workflow, vercel/chat, vercel/eve