Videos: agents as an OS for work

AI agents are being described less as chat windows and more as the layer that sits between workers and the software they use. (youtube.com 1) (youtube.com 2) In Tommy Geoco’s April 2026 video, “AI Designed This While I Was Sleeping,” he calls the category “agentic design” and says the “harness is air traffic control for your AI.” (youtube.com) The video’s chapter list says that harness has “three jobs” the model does not do, and ties the change to design tools, code handoffs, and a coming period of consolidation around Figma and MCP. (youtube.com) The other video, “Claude Just Shipped an Operating System. Most Enterprises Aren’t Ready,” argues Anthropic’s recent product moves add up to an operating layer, not a single assistant feature. (youtube.com) That framing lines up with Anthropic’s Model Context Protocol, or MCP, which the company introduced in November 2024 as an open standard for connecting AI assistants to business tools, repositories, and development environments. (anthropic.com) Anthropic’s documentation says MCP gives agents three basic building blocks: tools for actions, resources for data, and prompts for predefined workflows. It also separates local servers on a device from remote servers on the internet. (claude.com) (anthropic.skilljar.com) That matters because an agent that can open files, search Slack, query GitHub, or change a record in a work app needs the controls people expect from software, not just a chatbot. Anthropic says users authenticate each connector separately and that permissions mirror the user’s access on the outside service. (claude.com) Anthropic’s API docs also show why engineers are talking about orchestration. The company’s MCP connector lets one Claude request connect to multiple remote servers, choose enabled tools, and pass OAuth tokens for authenticated systems. (platform.claude.com) The same pattern is showing up outside Anthropic. GitHub said its Agent HQ effort is about orchestrating “a fleet of specialized agents,” while Amazon Bedrock AgentCore describes the orchestration layer as the loop that calls the model, picks tools, manages context, and handles failures. (github.blog) (aws.amazon.com) The design version of that shift is that interfaces may move from fixed screens toward systems that generate work products, call tools, and hand results back to people for approval. Geoco’s video says that changes the round-trip between design and code and predicts “headless design” within 18 months. (youtube.com) Anthropic’s connector ecosystem is also expanding beyond office software. Reporting this week said Claude added new connectors for services including Spotify, Uber, Instacart, and tax tools, extending the same model of delegated actions into consumer apps. (9to5google.com) (9to5mac.com) The common claim in both videos is that the hard problem is moving away from model intelligence alone and toward the system around the model: what it can touch, what it should remember, what it is allowed to do, and how a human checks the work. (youtube.com 1) (youtube.com 2)