Welcome back. Anthropic’s week isn’t going according to plan. Following reports that Claude Code has actually regressed lately, OpenAI stole the spotlight by dropping a new model that makes Opus 4.7 look expensive in comparison.

Also: Amazon engineer's guide to pivot from SWE to AI engineer in 2026, 12 Agent Skills from Google, and why a top analyst predicts mass protests against Anthropic and OpenAI.

OpenAI debuts its most agentic model yet: The ChatGPT maker just dropped GPT-5.5, their smartest model to date. It's designed to handle entire workflows from start to finish rather than just one-off tasks. It excels at coding, debugging, and managing data across various tools, all while being much more token-efficient than previous versions in Codex. Nvidia has already deployed it to over 10,000 employees. Watch GPT-5.5 in action.

DeepSeek drops V4 preview tuned for agent workflows: The Chinese AI startup just previewed its highly anticipated next-gen LLM, available in open-source pro and flash versions. It shows strong performance in agent-based tasks, knowledge processing, and inference, with lower costs than earlier releases. The model supports agent tools like Claude Code and runs natively on Huawei's Ascend processors. Developers can find the open weights on Hugging Face.

Claude agents can now remember across sessions: The AI lab just launched a public beta for built-in memory on its Managed Agents platform. This allows long-running agents to maintain context across sessions rather than starting from scratch every time. Memory is stored on a filesystem, so agents can use the same bash and code tools to read and write their own notes. Teams get audit logs, scoped permissions, and shared stores, with all API-tracked changes being reversible.

A dental ed co runs its backend on Viktor: crons, reports, automations. A coaching firm sends midnight tasks; Viktor finishes them by morning. A 4x4 parts co chews its backlog. A landscaping startup built 63 workflows in 2 weeks. A DTC brand runs 11 use cases, from checkout debugging to Klaviyo audits. A tutoring co runs DB enrichment and email campaigns.

One coworker. Every department. Lives in Slack, connects to 3,000+ tools, does real work. 9,000+ teams. SOC 2.

Get started for free →

Pilots without payoff. Two years into the AI boom, most companies still can't show their boards any real ROI. McKinsey calls this the "great AI paradox." Despite 80% of firms investing heavily, there's little to show for it. Senior partner Alexis Krivkovich points out that while the models work, companies fail because they're just bolting new tech onto old, human-centric workflows.

End to end. Krivkovich highlights McKinsey's work with the American Arbitration Association, where AI agents now handle the grunt work. Agents build timelines, analyze cases, and draft summaries. With AI doing the prep, humans move "above the loop" to focus entirely on review. In this new workflow, judgment becomes the entire job.

The billion-dollar catch. Flattening workflows often leads to an obvious next move, which is to stop hiring juniors. Many companies already have. But junior tasks are where future managers build their judgment. Krivkovich calls this the billion-dollar question. Cut early-career roles now, and in five years, you'll be poaching senior talent at a massive premium.

The real work. Building an agentic organization goes beyond tech. It requires rethinking roles, reviewing processes, and leadership training. In this model, change management stops being a project and becomes the new status quo.

Stop fighting with CAPTCHAs and Geo-blocks when building agents.

Instead of stitching together proxies, browsers, and scrapers, Bright Data’s MCP Server gives AI agents real-time, unrestricted access to the web—Claude Desktop, Cursor, Windsurf, and more.

Just make the request, it handles the rest.

Try Bright Data’s MCP at no cost.

Codex sends your prompt directly to the model, so accidentally pasting secrets like API keys means they could leave your machine. You can lower that risk by adding a pre-submit hook to scan your prompt and block it if a key is detected.

Save this in your Codex config (format may vary depending on your setup):

This checks for common OpenAI, GitHub, and AWS key patterns before your prompt is sent. If a match is found, the request is stopped.

How to run a self-hosted AI agent on your machine: This tutorial shows developers how to set up and use Hermes Agent, a self-improving AI tool. You'll learn to deploy it on a virtual server, connect it to Telegram, and add features like web search, voice commands, and automated daily scheduling.

OpenCode: An open-source AI coding agent that runs in your terminal, IDE, or desktop to help you write, understand, and refactor code using your favorite models.

Awesome Agent Skills (18.5k ⭐): A curated library of over 1,000 agent skills from official developers and the community, fully compatible with Claude Code, Codex, Gemini CLI, Cursor, and more.

How far are we from agents that can self-generate world knowledge? Most AI agents stall out because they rely on constant human feedback and rewards. However, teaching agents to explore new environments on their own allows even smaller models to beat out major systems like Gemini without needing a hand-held approach.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 240K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team