Gemini computer use and AI SDK updates #59
Today's Letter
- Google DeepMind, Gemini 3.5 Flash computer use introduced
- Vercel, AI SDK 7 release
- OpenAI and Broadcom unveil LLM inference chip
- GitHub adds self-service credential revocation
- Cloudflare Workflows adds saga rollback handlers
Google DeepMind, Gemini 3.5 Flash computer use introduced

- Google DeepMind introduced computer use support in Gemini 3.5 Flash in a post dated 2026-06-24
- The announcement positions computer use as a new capability for Gemini 3.5 Flash rather than a separate standalone product line
- Related model references in the source materials include Gemini 2.5 computer use model, indicating a broader computer-use model track inside Gemini
- The release is relevant to agent workflows that need browser or desktop-style task execution through Gemini APIs and tool-driven orchestration
- Google published the announcement through its official channels, which supports treating the launch as a confirmed product update
- The provided source excerpt does not include pricing, quota, benchmark, or availability details, so those fields are not stated here
Source: blog.google
More: androidauthority.com · sqmagazine.co.uk · fonearena.com
Vercel, AI SDK 7 release
- Vercel released AI SDK 7 on June 25, 2026.
- The TypeScript SDK reports more than 16 million weekly downloads.
- The update groups changes across five areas: development, runtime, integration, observability, and multimodal support.
- Reasoning control is now exposed through a unified `reasoning` option for `generateText` and `streamText`.
- Typed tool context and runtime context were added for tool inputs, step preparation, and approval flows.
- A new `uploadFile` API uploads files once and reuses provider references across later model calls.
- Runtime features include tool approvals, `WorkflowAgent` durability, timeouts, and sandbox support.
- AI SDK 7 also adds support for external agent harnesses such as Codex, Claude Code, Deep Agents, OpenCode, and Pi.
- Vercel provides `npx @ai-sdk/codemod v7` for migration from AI SDK 6.
Source: vercel.com
More: blog.luden.io
OpenAI and Broadcom unveil LLM inference chip

- OpenAI and Broadcom unveiled an LLM inference-optimized chip.
- The confirmed scope is inference optimization for large language models.
- The provided source material identifies OpenAI and Broadcom as the two companies tied to the announcement.
- No model names, performance figures, pricing, or release dates are included in the supplied materials.
- Additional technical specifications and deployment details were not confirmed in the verified inputs.
Source: techcrunch.com
More: engadget.com · news.hada.io · letsdatascience.com
GitHub adds self-service credential revocation
- GitHub added break-glass actions for GitHub Enterprise incident response on June 24, 2026.
- Enterprise owners and members with Manage enterprise credentials can revoke SSO authorizations.
- Targets include personal access tokens, SSH keys, and OAuth tokens across the enterprise.
- The same roles can delete user tokens and SSH keys across the enterprise for EMU accounts only.
- Organization-level listing and revocation of SSO authorizations are available only through REST APIs.
- Enterprise members now get a Settings -> Credentials view with counts of generated or SSO-authorized credentials.
- Users can revoke or delete all of their credentials and authorizations in one action.
- Revoked and deleted credential details are recorded in audit logs and email notifications.
Source: github.blog
Cloudflare Workflows adds saga rollback handlers
- Cloudflare introduced saga rollbacks for Cloudflare Workflows on 2026-06-25, adding per-step compensation logic for multi-step durable workflows
- Developers can now pass a `rollback` function in the options object of `step.do()`, instead of managing reversal logic in a shared `try/catch` block
- Rollback handlers are intended for cases where a failed workflow leaves earlier external operations in a partial state, such as debits, credits, inventory reservations, or shipment creation
- If a workflow fails terminally, eligible rollback handlers run in reverse step-start order rather than reverse completion order, including the failing step if it registered a rollback handler
- Cloudflare said rollback functions should be idempotent, similar to normal workflow steps, with provider idempotency keys recommended for both forward and reverse operations
- The API is designed to handle partial execution cases where a step may have affected an external system before failing, so rollback handlers must tolerate `output === undefined`
- Cloudflare rejected a fluent `step.do(...).rollback(...)` design during development because `step.do()` already returns a Promise-like durable step output and interacts with Workers RPC promise pipelining
- The feature reduces custom compensation bookkeeping for long-running workflow orchestration while preserving Workflow durability for both normal execution and rollback paths
Source: blog.cloudflare.com
Jocoletter curates AI, software, and product trends for developers and builders.
#Cloudflare #GitHub #Google #OpenAI #Vercel