Vercel explains how its agent handles permissions on live systems: it first investigates a problem, then proposes a plan describing the actions it intends to take. The approach tries to avoid both overly broad access and a flood of approval prompts that users end up clicking through.
Vercel's MCP server now has a submit_feedback tool that lets an agent send bug reports, friction reports and feature requests to the Vercel team, with structured context about the task. It requires sign-in and only sends feedback when the user asks for it.
As AI agents move beyond generating text or analysing data, 1Password is introducing controls over how they use credentials.
Vercel Connect can now create a separate, restricted Stripe API key for each agent or task, limited to the permissions the user selects. The aim is to stop agents from running on Stripe's default key, which grants full access to an account.
Version 1.56.0 of the open-source Goose agent can set the model's thinking level automatically and resume subagents that failed or were cancelled. It also adds a cap on concurrent subagents, a Windows installer for the desktop app and a new provider, along with bug fixes.
Analysts argue that fears of AI agents upending the banking industry are exaggerated, at least for the moment.