The latest articles on DEV Community by Boyu Du (@dbellworkshop). https://dev.to/dbellworkshop https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3846992%2F60e75915-6b65-408a-b59b-8f99e52f528d.jpeg https://dev.to/dbellworkshop en Boyu Du Sat, 28 Mar 2026 04:22:05 +0000 https://dev.to/dbellworkshop/why-im-building-agentfleet-from-one-assistant-to-a-visible-ai-team-3d6h https://dev.to/dbellworkshop/why-im-building-agentfleet-from-one-assistant-to-a-visible-ai-team-3d6h <p>Most AI tools still feel like hiring one very fast assistant.</p> <p>That is useful, but it still breaks down for real work.</p> <p>When I am working on something messy, I usually do not want one system to answer everything in a single thread. I want something closer to a team:</p> <ul> <li>one part researching</li> <li>one part planning</li> <li>one part executing</li> <li>clear checkpoints where the human steps in only when it matters</li> </ul> <p>That idea is what pushed me to start building AgentFleet.</p> <h2> What I am trying to build </h2> <p>AgentFleet is an open-source multi-agent workspace where the workflow is visible.</p> <p>Instead of hiding everything behind one chat box, I want the operator to be able to see:</p> <ul> <li>which agent is doing what</li> <li>where work is blocked</li> <li>when coordination is happening</li> <li>when a human decision is actually needed</li> </ul> <p>My bet is that AI products become much more useful when they feel less like "one assistant" and more like "a visible team with roles and checkpoints."</p> <h2> Why the visible office metaphor </h2> <p>I know the obvious risk: a visual metaphor can help people understand a workflow, but it can also feel gimmicky.</p> <p>I chose the office/team metaphor because I wanted the structure of the system to be legible at a glance. If there are researchers, planners, and executors, I want people to see that immediately instead of inferring it from hidden state.</p> <p>I am still testing whether that tradeoff is actually worth it.</p> <h2> Current state </h2> <p>This is still early.</p> <p>I am also in the middle of a rename from AgentsOffice / AgentOffice to AgentFleet, so some docs and screenshots still show the old name while I clean things up.</p> <p>The repo is here:</p> <p><a href="https://github.com/DBell-workshop/AgentFleet" rel="noopener noreferrer">https://github.com/DBell-workshop/AgentFleet</a></p> <h2> Feedback I would really value </h2> <p>If you have built agent workflows, internal tools, or operator-facing AI products, I would love your blunt take on two things:</p> <ol> <li>Does a visible multi-agent workspace solve a real usability problem?</li> <li>Does the office/team metaphor help, or does it add unnecessary UI ceremony?</li> </ol> <p>I am especially interested in feedback from people who have already felt the limits of a single-threaded AI assistant.</p> ai productivity opensource showdev