PLATFORM
Multi-agentSandboxTiered memorySkills+ToolsUnified governance
Challenge: Business systems are fragmented, models and tools are hard to reuse, and there is no unified governance or audit trail—which makes scaling agents slow and expensive.
Architecture: Built on the crusbro-agent SDK with Supervisor/Worker multi-agent collaboration, sandboxed execution, tiered memory and the Skills/Tools protocol. Domain capabilities plug in as plugins, keeping business logic decoupled from models, storage and deployment.
Value: Agent launch time drops from ~8 weeks to ~2 weeks; tools, models and prompts upgrade independently; every step is covered by an auditable, replayable evidence chain.
"Developer interfaces, runtime governance and a safe sandbox come together in one SDK—our team can focus on the domain." — Platform architect
−75%agent launch time
100%replayable tool calls
4 tiersof memory
ROBOT AGENT
Challenge: Business apps call drivers directly, leaving no auditable boundary between the LLM and safety controls.
Architecture: Application → crusbro robot-agent → Skills/Tools → ROS2 capabilities / HAL; high-risk actions route through human-in-the-loop gates.
Value: Plans are replayable; high-risk actions require dual review; chassis and sensors become swappable through the HAL.
"Robots and agents now share the same audit trail." — Robotics tech lead
OPC AGENT
Challenge: Many processes, many hand-offs, little traceability—a single operator struggles to balance throughput and proof.
Architecture: The One-Person Company (OPC) Agent combines workflow orchestration with cross-system Skills and automatically consolidates execution evidence.
Value: ~50% faster end-to-end cycle time; cross-system actions stay replayable.
"One person can run a full delivery loop end to end." — Independent consultant
FREIGHT AGENT
Challenge: Booking, customs, trucking and loading data is scattered across inboxes and spreadsheets; anomaly triage is slow.
Architecture: The Freight Agent orchestrates booking, customs, trucking and loading Skills with freeze-to-send approval and end-to-end evidence.
Value: ~58% faster anomaly triage; every step is auditable.
"What the approval preview shows is exactly what gets sent." — Head of operations
DRAWING MODEL
Challenge: Mixed native CAD and scans; 3D assemblies out of sync with 2D drawings; BOM repeatedly realigned between design and ERP; routing re-keyed with slow change propagation.
Architecture: Drawing foundation model (3D/2D + BOM) → review workbench (BOM diff, routing draft) → MES / process-route APIs; model version, primitive and field lineage land in the audit store.
Timeline: 8-week pilot, 16-week line rollout (incl. data governance).
ROI: ~74% less process-prep labor; ~31% fewer monthly rework tickets (same-batch baseline).
"We can open any field back to the source graphic—far fewer arguments in meetings." — Process manager
OPC × FREIGHT
Challenge: Night-shift staff re-handle the same emails; booking and customs data scattered across mailboxes.
Architecture: Multi-agent stack + freeze-to-send approval queue; project-scoped file vault.
Timeline: 6 weeks joint integration, 4 weeks pilot.
ROI: −2 clerical FTE (night shift); ~58% faster anomaly triage.
"The approval preview is exactly what goes out." — Head of operations
VISION
Challenge: High SKU churn; rule-based vision is expensive to maintain.
Architecture: Edge inference + central training queue; defect taxonomies hot-updated.
ROI: Miss rate 1.2% → 0.35%; ~35 cold-start samples per new SKU.
"We cut over the same night as the line change." — Quality manager
