Sending a task from my phone is worthwhile only if the agent does its share of finding context. If I have to prepare a dossier on a tiny screen, the convenience is gone before work begins. I want to send a clear reference and get on with my day while the task moves within that request.
Dispatch appeared in Anthropic's March 2026 computer use announcement, with the promise of starting tasks from a phone and executing them on a computer. Revisiting the news in October 2026, I care about continuity between those screens. Sending a message away from my desk is easy to understand. What happens afterward determines its value.
In September 2026, I expected an agent to find the information it needed before sending questions back to me. In June, I had already objected to passing the context of every job to other agents. Both expectations fit the same request: research what you need without turning every assignment into a copy of my entire working life.
Imagine starting an investigation from a phone into a change already described in the project. That is an example I want to test. The agent has a reference, locates the description, and reads the related material. Asking again for information already written there just transfers research to the person typing with a thumb. Excellent use of automation.
A question about my intention has a different value. If two projects are plausible and there is no sufficient clue, a short question helps. I want that distinction in practice: the agent should recognize when my decision is missing and when it simply needs to open a document. Being brief with me requires doing some work before answering.
Sending all context everywhere also bothers me. Decisions from another job can end up in a task where they have no business being. I prefer relevant references and a search guided by the request. I need to understand why information entered the execution, especially when the conversation starts away from the computer.
Back at my desk, I want the result attached to the request I sent. I also want to see what remains without reconstructing an entire conversation. My next criterion for this kind of tool is a short request with accessible supporting material: watch what it finds independently and which questions come back. The convenience needs to survive the first tap on send.