Why it mattersThe review names four design choices that work. A team building its own harness can read them and compare them with its own product.
A developer tries a new personal-assistant harness every month. Most of them ask for every password on day one. One of them stops asking, and a working review is now circulating. Nathan, the writer behind fig.blue, posted a review of Claude Duet on 29 October. The title is "Duet is the future of chatbot interaction". The post reached 124 points on Hacker News. Nathan says he wrote it himself. He adds that he is not working with Anthropic.
Claude Duet is Anthropic's longer-running chat harness, released on 21 October. Nathan says Duet is the first harness of its kind that feels useful rather than tedious. His review names four design choices and says the mix is what works. The piece is a user's account, not a benchmark. Nathan supports each point from his own use.
Fast replies, then background work
Nathan says Duet gives a short answer first and then goes to work on the longer task. He writes that it "clearly prioritizes getting a snappy response to the user even if background work is needed". The model behind it is Anthropic's Opus 5.5. Nathan says the harness appears to be tuned for how fast the first reply lands.
Channels, not one long thread
The second point is organisation. Duet opens a new channel for each task, and uses Command-K to move between them, the same way Slack does. Nathan says this stops a long chat from losing older work. "Duet does the tedious session organizing for me," he writes.
No credentials asked for
The third point is the one Nathan says matters most. Duet does not ask for a username and password to any service. Each channel can be limited to a single GitHub repository. Nathan writes that Duet "stores no credentials and actively discourages being given them". He says the harness uses the connector ecosystem that already exists instead of asking the user to hand over keys.
Short answers, with a small signal of life
The fourth point is tone. Duet gives short answers to short questions. It uses an emoji reaction to say it has read a message. The user then knows the assistant is still there. "Nobody wants to read a full page response to a simple question," Nathan writes.
What he says is still rough
Nathan is clear that Duet is in beta. He names visible threading problems. He also names access handling that sometimes fails to attach to a repository. He compares Duet with Muse, Dot, and Hermes. He notes voice input is still better in Dot. He adds that this is one person's week with one product. His review is a report of his own use. It is not a measurement anyone can repeat yet.
Why a Hacker News discussion of a product review matters
The 124 points on the discussion thread come from readers of a working developer's review. The comments do not argue with Nathan's claims about Duet. They argue about whether a chat harness that asks for no credentials can do real work on your behalf. That tension, between useful and safe, is the hard problem. Every team releasing an agent to an everyday user is working on it today.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
