AI wants more than a text box
Devin’s tools, Claude’s computer controls and MCP’s integration standard pose a practical question: what are you ready to delegate?
Cognition’s Devin announcement presented a software agent equipped with a shell, code editor and browser in a sandboxed environment. The company showed examples of building applications and fixing code, and described a system that reports progress and accepts feedback. The proposition was broader than another suggested line of code: assign a task, then review what the system produces. These were Cognition’s demonstrations and claims, with their specific settings attached. [1]
Anthropic’s Claude computer-use release approached existing software through screenshots, cursor movement and virtual keyboard input. The company described a public beta that could turn instructions into computer actions, while acknowledging slow and error-prone behavior. It also identified prompt injection as a concern when a model encounters malicious instructions on screen. That makes the interface itself part of the story: the same screen that presents useful information can present instructions the user never gave. [2]
MCP addressed a different layer. Anthropic introduced it as an open standard for connecting assistants with data sources and tools through clients and servers. The company named early integrations and supplied example servers. Its stated aim was to reduce the need for a separate custom integration for every source. This is connective infrastructure. My interpretation is that easier access to context can support more useful work, while leaving the quality of the resulting decisions to be tested. [3]
The performance claims need their settings attached. Cognition reported resolving 13.86% of issues on a random quarter of SWE-bench, with differing assistance conditions for the comparison models. Anthropic reported Claude’s computer-use result on OSWorld. These evaluate different kinds of work; they cannot be turned into a single league table of autonomy. A useful year-end question is what each result actually permits a reader to conclude. [1] [2]
Cognition’s launch invitation was for early access; Anthropic described Claude’s ability to follow instructions through computer input. Those offers are useful evidence about what the companies were building, not a census of successful deployments. My editorial reading is that delegation was becoming more concrete. I find that encouraging for my long-term ambitions. For yours, start with a task you can evaluate and a result you can inspect. [1] [2]
