AI agents move into your inbox
Your inbox just became an agent interface. Today’s updates make it easier to hand off email work, test coding models, and trace the apps you ship with them.
In today’s edition
- Claude can now send Gmail replies and manage Google Drive files on paid plans.
- Perplexity Computer can run tasks from an email thread, with an audit trail.
- GLM-5.3 is live on OpenRouter with large reported gains on coding-agent benchmarks.
- Vercel is putting $1 million behind testing its sandbox boundaries.
- Supabase now carries a trace ID from the browser into its logs.
Claude can now act in Gmail and Drive
Claude can now draft and send Gmail replies and manage files in Google Drive. You decide when an action needs your approval.
- Connect Gmail or Google Drive from Claude’s connectors menu.
- The capability is available on all paid plans.
- A simple first test: hand Claude a recurring inbox task, such as drafting responses to common customer questions, then keep approval turned on.
- Use Drive access for workflows that need to find, organize, or work from files before producing a response.
What it means Start moving repeatable inbox and file chores into approval-based Claude workflows, but keep a human checkpoint on anything customer-facing, financial, or irreversible.
Source: claudeai
Perplexity turns email into an agent trigger
Perplexity Computer now works through email: send, forward, or cc computer@perplexity.com on a thread to start a task.
- Each email task runs as a normal Computer session.
- You can view those sessions on web and mobile.
- Tasks keep the same audit trail as work started inside the app.
- The feature is available to all Computer users today.
What it means Try email as the front door for delegated work: forward a request to Computer when you want an auditable task without opening another app.
Source: perplexity_ai
GLM-5.3 lands on OpenRouter
GLM-5.3 from Z.ai is now live on OpenRouter, using the same base model as GLM-5.2 with gains attributed entirely to post-training.
- OpenRouter reports Terminal-Bench 3.0 rising from 4.6 to 28.3.
- It reports DeepSWE v1.1 rising from 46.2 to 66.9.
- The release gives builders another model option for terminal and software-engineering tasks.
- Because the base model is unchanged, the update is a useful reminder that post-training can materially change tool-use performance.
What it means Add GLM-5.3 to your coding-agent eval set now, using your own tasks and failure cases rather than choosing a model from benchmark scores alone.
Source: OpenRouter
Vercel tests agent sandbox defenses in public
Vercel says agents can exploit vulnerable sandbox boundaries, so it is testing Vercel Sandbox in the open with a $1 million hacker challenge.
- The challenge targets escape from the Firecracker microVM.
- It also targets the host-side network boundary.
- Vercel is offering up to $50,000 per report through HackerOne.
- The announcement puts agent runtime isolation, not just prompt safety, in focus.
What it means Treat sandbox isolation and outbound network controls as product requirements for agent features, especially when your agents run code or handle customer data.
Source: vercel
Supabase joins browser and backend logs
Supabase added trace propagation so one trace_id can run from your browser into Supabase logs, replacing timestamp guesswork when you chase a broken request.
- The feature is built on the open W3C Trace Context standard.
- It works with OpenTelemetry, Sentry, Datadog, Honeycomb, and Grafana.
- It is available in supabase-js, Swift, Flutter, and Python.
- Supabase also refreshed Log Drains documentation and added two vendors.
What it means Turn on trace propagation before your AI app gets harder to debug, so you can follow a user request from client behavior to backend logs.
Source: supabase