AI engineering, November 14, 2026

Top News

ChatGPT Wallet is a way for agents to complete purchases, with user approval and existing card protections. Agents can use it in ChatGPT Work, Codex, and dots.

Here's how it fits into an agent workflow:

• Agents can complete purchases through the wallet. User approval is part of that process, rather than leaving the purchase entirely to the agent.
• ChatGPT Work, Codex, and dots are the places agents can use it.
• Existing card protections apply to these purchases, so using an agent does not mean setting those protections aside.

One catch: purchases require user approval. An agent can handle the purchase, but your workflow still needs the user's decision before it goes through.

Sources: openai.com

Top News

Running an agent browser does not have to mean running a browser container. Cloudflare has open-sourced Kitesurf, its browser engine for AI agents, so you can inspect the code, run it locally, or deploy an instance to your own Cloudflare account.

Here's what you can run:

• The engine renders web pages on Cloudflare Workers without a browser container.
• CDP support lets you use clients such as Puppeteer to control the engine.
• The GitHub source is available under Apache 2.0, so you can inspect the implementation before hosting it yourself.

One catch: Kitesurf is experimental, with standards support, rendering fidelity, and performance still under development. A self-hosted instance also needs an external authentication layer, such as Cloudflare Access.

Try it: Follow the build and setup instructions, then run pnpm run dev to build the engine and start Wrangler's development server.

Sources: developers.cloudflare.com

Top News

Amazon Bedrock AgentCore is AWS's managed runtime for deploying agents without maintaining the underlying infrastructure. The updated runtime is available today in AWS Regions across North America, South America, Europe, and Asia Pacific.

Here's what changed:

• Sessions start up to 15× faster, with up to 4× more compute than the earlier runtime.
• Memory is reclaimed after an agent releases it, and idle agents scale to zero. Multi-turn agents with natural pauses consume up to 68% less memory per session.
• Committed baseline pricing bills a per-session memory minimum, with on-demand burst capacity up to four times that baseline.
• Sessions support up to 8 vCPUs and 32 GB, with starts under 2 seconds for container images up to 10 GB.
• Each session runs in its own hardware-isolated Firecracker MicroVM with an identity, access policy, and audit trail.

One catch: session-labeled credentials for per-session storage scoping are planned by re:Invent, not available yet.

Sources: aws.amazon.com

1. vLLM runs Inkling 2 Small at up to 630 tokens/second per user on GB200 with TP4
2. Google's Agent Anomaly Detection flags 87% of executed harmful instructions for post-execution review
3. Nebius plans $40 billion in AI infrastructure spending for 2027, funded by customer prepayments and loans · Reported · techcrunch.com
4. OpenAI ships Codex Security CLI and TypeScript SDK to find and fix vulnerabilities · 29 Stars today
5. Nenya claims its 0.5B-active MoE rivals Qwen3.5-2B on pretraining benchmarks using 500× less training compute
6. PhysLoop uses LLMs to repair physics errors in AI-generated videos at inference time · 42 Upvotes