Agents gain their own cloud computers, payment rails and a visible working surface
SpaceXAI runs persistent agents on dedicated cloud computers through Grok Bot; AWS opened a budget-capped payment layer, and GitHub Copilot added a visible working surface that tracks agent progress.
Artificial Intelligence··Night
Grok Bot runs persistent agents on their own cloud computers
Grok Bot, introduced by SpaceXAI, runs persistent agents that connect to websites, applications and inboxes and carry multi-step work through from start to finish. As InfoQ reports, the agents keep conversation context and user preferences, save processes they learn by watching and re-run them, work in parallel and talk to one another, and ask for human approval when a decision is needed. The company separates the system from coding-focused agents such as Claude Code and OpenAI Codex: Grok Bot targets broad business workflows including sales, marketing and operations, and it grew out of a prototype used inside the company. The beta is open to SuperGrok Heavy and to Cursor Ultra and Cursor Teams Premium subscribers; the announcement came right after SpaceX completed its acquisition of Cursor.[1]
AWS opens a payment layer that caps agent spending
Amazon Bedrock AgentCore payments lets AI agents reach paid APIs, paid data feeds and metered services through wallet integration, spending limits and a shared payment layer. According to the AWS Machine Learning Blog, spending limits are enforced through recipient allowlists, per-payment ceilings, session budgets and expiry dates; the layer supports HTTP-native payment protocols including x402 v2 and the Machine Payments Protocol. The sample build runs x402 v2 on the Base Sepolia test network, and the post says Coinbase CDP and Stripe Privy wallets, Ethereum, other EVM-compatible chains and Solana are supported. The AWS post does not state general availability or pricing, so the announcement should be read as a preview layer rather than a shipped general-release product.[2]
GitHub Copilot gathers agent workflows on a visible surface
GitHub's canvases, a durable shared surface, keep an agent's progress, decisions and approval points in one inspectable place instead of a chat stream; the feature is available in the GitHub Copilot app through the /create-canvas command. A post by developer advocate Ayan Gupta on The GitHub Blog cites migration projects, content creation and multi-step orchestration as examples; the pattern it proposes covers defining workflow states, surfacing important decisions, writing progress immediately and leaving human approval points explicit. The post also gives the AI credits two reference builds consumed: about 3,000 for Java Modernization Studio and about 2,000 for Site Studio; Gupta argues that this upfront cost reduces repeated prompting and rework.[3]
