Argus

Inspiration

Every company deploying AI agents is flying blind. There's no visibility into what agents are doing, what they cost, or when they need human intervention. We wanted to build the missing control layer — Datadog for AI workforces.

What it does

Nexus is a real-time 3D command center for AI agent fleets. Type one prompt, and Nexus autonomously decomposes it into tasks, spawns specialized agents across 8 departments, and streams live output — code, designs, documents, SVG graphics, interactive dashboards — all viewable in a built-in asset browser. Human-in-the-loop escalations, real-time cost tracking, and fleet-wide orchestration from a single interface.

How we built it

React Three Fiber for the 3D scene, Express + Socket.io for real-time WebSocket communication, Zustand for state management, and the Claude API powering every agent. The Nexus Brain uses Claude to decompose prompts into agent tasks. We built scaling infrastructure (batched updates, selective output subscriptions, shared geometries) that handles 200+ concurrent agents. All TypeScript, all from scratch in 48 hours.

Challenges we ran into

Zustand selectors returning new array references on every render caused infinite re-render loops. Socket.io broadcasting per-character output from 50+ agents simultaneously melted bandwidth — we solved it with output throttling and selective subscriptions. Fitting the Vite + Three.js build into Railway's memory limits required stripping 40% of the bundle.

Accomplishments that we're proud of

One prompt spawns 50+ real Claude agents that produce actual code, SVG graphics, HTML websites, pitch decks, and dashboards — all viewable live in the asset browser. The 45-second auto-pilot demo runs hands-free with cinematic camera work, narration, and a mid-demo fleet pivot that redirects every agent in real-time.

What we learned

AI orchestration is a UX problem, not just an infrastructure problem. Watching agents work in 3D with particle effects and status rings makes fleet management intuitive in a way that logs and dashboards never could. Also: scaling from 15 to 200 agents is an entirely different engineering challenge than building for 15.

What's next for Argus

Agent-to-agent communication — letting agents pass outputs to each other in pipelines. Persistent memory so agents learn from previous runs. Multi-model support beyond Claude. And turning the asset browser into a full IDE where you can edit, version, and deploy agent-generated code directly.

Built With

• anthropic
• claude
• node.js
• vite
• webspatial