[{"data":1,"prerenderedAt":129},["ShallowReactive",2],{"blog-\u002Fblog\u002Fwhy-three-engineer-agents":3},{"id":4,"title":5,"author":6,"body":7,"cover":114,"description":13,"draft":115,"extension":116,"meta":117,"navigation":118,"path":119,"publishedAt":120,"seo":121,"stem":122,"summary":123,"tags":124,"updatedAt":127,"__hash__":128},"blog\u002Fblog\u002Fwhy-three-engineer-agents.md","Multi-Agent AI Coding: Why Veflow Uses Three Specialists","Veflow team",{"type":8,"value":9,"toc":106},"minimark",[10,14,37,40,45,52,55,59,62,68,79,85,95,99],[11,12,13],"p",{},"Multi-agent AI coding is everywhere in 2026, but most teams still route every task to one anonymous generalist and wonder why the results feel flat. Veflow takes a different approach: assign a task and the picker offers three named engineer agents.",[15,16,17,25,31],"ul",{},[18,19,20,24],"li",{},[21,22,23],"strong",{},"Atlas"," — the full-stack generalist: UI, API endpoints, tests, refactors, migrations.",[18,26,27,30],{},[21,28,29],{},"Nova"," — backend & systems: Postgres, auth, background jobs, caching, rate limiting.",[18,32,33,36],{},[21,34,35],{},"Orion"," — API & infrastructure: REST and GraphQL design, CI\u002FCD, observability, integrations.",[11,38,39],{},"A fair question is why. One sufficiently capable model could take any of these tasks — so is a roster of three just theater?",[41,42,44],"h2",{"id":43},"the-honest-version-of-what-differs","The honest version of what differs",[11,46,47,48,51],{},"Let's not oversell it. Atlas, Nova, and Orion run on the same model and share the same execution machinery — the same isolated environment, the same tools for reading, editing, running commands, and opening PRs. What differs is ",[21,49,50],{},"the system prompt",": each agent carries its own role, its own capabilities, its own framing of what it's for. The specialisation is in the briefing, not in a different brain.",[11,53,54],{},"If that sounds underwhelming, consider that it's exactly how human specialisation works on small teams. Your backend person and your frontend person have the same IDE, the same shell, the same git. What makes them specialists is focus — which concerns they treat as central versus peripheral. Prompting is the agent version of focus, and the difference shows up where it matters: a backend-framed agent reaches for the migration and the index; a full-stack-framed one thinks about the form state and the empty state.",[41,56,58],{"id":57},"why-a-multi-agent-roster-is-really-for-the-humans","Why a multi-agent roster is really for the humans",[11,60,61],{},"The less obvious payoff is on your side of the screen.",[11,63,64,67],{},[21,65,66],{},"Dispatch gets fast."," \"This is a Nova task\" is a quicker, more repeatable decision than re-deriving an ideal prompt framing per task. The lanes are a vocabulary.",[11,69,70,73,74,78],{},[21,71,72],{},"Plans become legible."," When Polaris breaks an idea into tasks, it routes each one to an agent by lane. A plan that reads ",[75,76,77],"em",{},"schema and jobs → Nova, endpoints → Orion, the feature surface → Atlas"," communicates its shape at a glance — the assignments are information.",[11,80,81,84],{},[21,82,83],{},"History reads better."," \"Nova handled the rate limiter\" carries more signal in the activity feed and the task history than \"the agent did it.\" Named actors make a multi-agent workspace narratable.",[86,87,92],"media",{"aspect":88,"hint":89,"kind":90,"src":91},"aspect-[16\u002F10]","Assign popover open on a board task, listing teammates and the agent roster — Atlas, Nova, Orion, Polaris — each with a role line","Screenshot","\u002Flanding\u002Fagents-picker.png",[11,93,94],{},"Named lanes. The picker is a routing decision, not a config screen.",[41,96,98],{"id":97},"no-black-box-either-direction","No black box, either direction",[11,100,101,102,105],{},"The flip side of \"it's the prompt\" is that there's nothing mystical to evaluate. Each agent's role, bio, and capabilities are visible in the workspace — what you read is what the agent is told it is. We'd rather be plain about that than imply each one is a bespoke model. The bet isn't that three prompts beat one model; it's that ",[21,103,104],{},"named, scoped lanes beat an anonymous generalist"," — for the agent's focus, and even more for the humans deciding, reading, and reviewing around it.",{"title":107,"searchDepth":108,"depth":108,"links":109},"",3,[110,112,113],{"id":43,"depth":111,"text":44},2,{"id":57,"depth":111,"text":58},{"id":97,"depth":111,"text":98},"\u002Fblog\u002Fwhy-three-engineer-agents.png",false,"md",{},true,"\u002Fblog\u002Fwhy-three-engineer-agents","2026-06-18",{"title":5,"description":13},"blog\u002Fwhy-three-engineer-agents","Veflow runs three named AI coding agents — Atlas, Nova, Orion — instead of one. Here's the honest case for specialized lanes over a single generalist agent.",[125,126],"product","philosophy",null,"HWgpl-i_Afm1h8VYVclzQFG6ifhdTuHoN7xMlGVWorU",1785474882190]