Band of Agents Hackathon · Track 1 · athenarocks

Deploy robots
without losing
the thread.

Orizon Fleet Coordinator is a deployment desk for our AMR fleet. Four agents — each a different framework, each a different open-source model on Featherless — coordinate in Band until the fleet config is safe enough to launch.

Try the live demoView source

The problem

Every client deployment is different.

We build AMRs and sell them B2B — clients buy fleets, not single units. Every factory has a different layout, size, and robot series. Today that handover is manual, step by step, and mistakes happen.

  • —B2B fleet sales — unique floor plan and config every time
  • —Days of manual site assessment, fleet config, and safety review
  • —Idle robots waiting for go-live — AMRs aren't cheap

What we built

A coordination room, not another dashboard.

Our engineer, on site, triggers a deployment. Four agents join a Band room, share structured JSON artifacts, and run a safety veto loop until ISO 3691-4 checks pass. A human approves the final package before Day 1.

Same product. Different floor every time. We're replacing the manual handover with a coordination room.

Four agents

Each owns one job. Each runs a different stack.

Cross-framework collaboration through Band — LangGraph, CrewAI, PydanticAI, and AnthropicAdapter on Featherless open-source models.

Site Assessor

Parses factory specs, flags Wi-Fi dead zones and ramp margins

LangGraphQwen2.5-32B

Fleet Configurator

Robot selection, zone assignment, charging rotation

CrewAI + RAGLlama-3.1-70B

Safety Reviewer

ISO 3691-4 checklist — holds veto authority

PydanticAIDeepSeek-R1

Launch Coordinator

Packages the deployment timeline and audit hash

AnthropicAdapterQwen2.5-72B

The workflow

Assess. Configure. Veto. Launch.

01

Assess

Site Assessor reads the factory spec and produces structured infrastructure flags.

02

Configure

Fleet Configurator picks robots from catalog, assigns zones and waypoints.

03

Review

Safety Reviewer runs ISO checks. Can veto and send the thread back.

04

Launch

Launch Coordinator ships the package. Ops Manager approves go-live.

Demo scenario

Acme Motors Plant 4 — 8 AMRs, automotive assembly.

First pass — veto

Site Assessor flags a Wi-Fi dead zone in Zone C. Fleet Configurator assigns 8× HeavyLift-X200. Safety Reviewer vetoes: robots in Zone C lose WMS connectivity — collision risk, no e-stop in the dead zone.

After revision — pass

Site Assessor adds a Wi-Fi 6 AP on column C-NE-7. Coverage goes from 95% to 99.2%. Safety Reviewer passes all ISO checks. Launch Coordinator packages a 3-day deployment timeline.

5–15 day manual process → ~10 minute coordinated agent session. The veto loop is the demo.

Fleet ops mode

Six robots. Real-time comms.

Switch to Fleet Ops in the dashboard. Drop a robot's battery below 20%, toggle an obstacle, change its zone — and watch the fleet coordinate in the chat room, powered by Featherless inference.

Open fleet simulator

Why this exists

A working idea I'm building toward.

I'm working toward Orizon Robotics — custom AMRs with FOC-controlled actuators. This project is a working prototype: a training exercise and mental model for how I want client deployments to flow — assess, configure, veto, launch — before the hardware is ready. I'm building the muscle memory now so the real ops layer is smoother later.

See the veto loop live.

Full 8-step workflow: assess → configure → veto → revise → pass → package → approve.

Launch demo