CRUSBRO · SHANGHAI · AGENT-OS

crusbro agent-OS Agent Operating System | Controllable delivery · 600k hours in the field

Our foundational product. Business automation, drawing, vision, robot and freight agents are built on it—stable, auditable, replayable.

  • Assemble: roles, workflows, rules, permissions, data sources — configure and run.
  • Control: HITL, freeze‑before‑send, end‑to‑end traceability, replay and human takeover.

Governance

Delivery and governance

From solution design and system integration to production operations, the crusbro agent-OS brings workflows, permissions, data boundaries and human review into one delivery framework—so agents run reliably, every action is traceable and every result can be verified.

Built for production

We configure, integrate and validate each agent around real business workflows before a staged launch.

  • Acceptance criteria tied to business outcomes
  • Phased rollout with minimal disruption

Transparent and traceable

Critical tasks, tool calls and human decisions are logged end to end, making exceptions easier to locate and results easier to review.

  • Auditable, replayable execution
  • Human review and takeover controls

Secure, sustainable operations

On-prem, VPC and edge deployment options combine with configurable access and data policies for reliable long-term operations.

  • Configurable data boundaries and permissions
  • Monitoring, optimization and operational support

Technology ecosystem integration examples

  • AWS
  • OpenAI
  • Anthropic
  • Qwen
  • MiniMax
  • Kimi
  • GLM
  • Baidu AI Cloud
  • Volkswagen

Industrial × AI

The combination of industry and AI is not about "smartness" but controllability

Real production isn’t an open world: system boundaries are clear, responsibility chains are clear, and error costs are calculable. We put AI where it belongs — as controlled execution units and structured extractors.

Control plane

Write "what can be done" as a switch and "when can it be done" as a gate.

  • The Tool gate control table is on by default and can be turned off one by one in the background.
  • Freeze → Preview → Send only after approval
  • Single-step replay and manual takeover

Data plane

Align model output to system fields and enable source location.

  • Field → Element/Page Number/Attachment Index Link
  • Project dimension file merging and status view
  • Low confidence fields go into manual queue

Integration

Treat "access cost" as the first constraint, rather than a patch after going online.

  • OPC UA subscriptions and controlled writes
  • IMAP/SMTP monitoring, classification and distribution
  • Contractual interface with MES/ERP/booking gate API
Audit
Sessions, tool calls, and outgoing queues are associated according to business order numbers; frozen snapshots can be reviewed each time they are sent.
Latency
Visual acceptance by workstation: delay, missed detection and replacement cost; decoupling edge reasoning and central training.
Boundary
The computer room/edge environment is divided into different environments; the outbound policy can disable the public network large model and only retain the local backend.

Chart

The risk falls from "semantics" to "interfaces"

The failure of industrial AI is usually not because the model is not smart enough, but because the boundaries are uncontrollable: outsourcing, writing, and version drift. We use "gate + freeze + replay" to convert risks into acceptable items.

Failure modes
Turn problems into monitorable interface events
UNCONTROLLED → CONTROLLED Outbound content freeze+approve device write guarded write Version drift pin+rollback Field traceability lineage

Chart

Delivery is not a feature list, it is an acceptance matrix

The same set of "controllable" methods falls on three core technology stacks: agents (actions and audits), drawings (fields and lineage), and vision (delay and type change).

Acceptance matrix
Each line has a signable indicator caliber
LINE AUDIT LATENCY BOUNDARY agent drawing Vision interface rollback freeze queue tenant lineage batch on-prem trace SLO edge

Technology stacks

Three Core Technology Stacks

Stable technology foundations support a focused portfolio of products and solutions. Each stack can be accepted independently or combined for a complete workflow.

Agent · Orchestration

Agent Architecture and Orchestration

SCENE · Freight Forwarding · Mail Document Chain · UA/Equipment

enterUA subscription, mailbox monitoring, business API callback.
deal withTool routing, RAG, manual gate; freight forwarding chain can be assigned according to box type.
outputControlled writing, pending queue, project file warehouse index.
Failure and acceptance criteria
failTool hallucination, accidental triggering, and confusion of session status → playback positioning.
acceptanceThe frozen snapshots are consistent, not rewritten after approval, and audited at the single number level.

Industrial Drawing Agent · Task orchestration

Industrial Drawing Agent

SCENE · 3D/2D DRAWINGS · PROCESS ROUTES · BOM

enter2D/3D drawings and CAD models: PDF / DWG / DXF / STEP / IGES / PRT and other mainstream formats, plus enterprise process standards and historical process cards.
deal withDrawing analysis → process-route generation → BOM validation → Excel/Word output → automatic drawing annotation; low-confidence results enter focused human review.
outputAuditable process routes, BOM and documents; field ↔ primitive/page/rule lineage; structured interfaces to PLM/MES/ERP.
Failure and acceptance criteria
failScan distortion, cross-page reference breaks, and symbol library inconsistencies.
acceptanceYou can click on the original image element for random inspection fields; you can align the field diff for change orders.

Inspection Agent · Device–Edge–Agent

Industrial Visual Inspection Agent

SCENE · AOI · Workstation re-inspection · End-edge agent collaboration

enterStation image streams, trigger signals, SKU metadata and inspection recipes.
deal withAgent orchestrates camera, lighting, model and rules by station; defect detection, dimensional measurement, positioning, assembly-compliance; low-confidence routed to HITL and alerts.
outputVerdict + disposition + evidence (original, parameters, model version, rule); QMS / MES writeback and replay.
Failure and acceptance criteria
failLighting drift, model change is not synchronized, and categories are unbalanced.
acceptanceSign the time delay, missed inspection, and model change window according to the work station; do not write "one number for the whole factory".

Field notes

From the field

What actually changed on site in three real delivery projects.

Equipment manufacturing · Drawing and process unificationDrawing key elements are extracted automatically; the agent generates process routes and a BOM, then exports results to Excel, Word and auto-annotated drawings. Engineers shift from re-typing to confirming and correcting, and any disputed field links straight back to the original annotation.
Parts maker, East China · Industrial Drawing Agent
International freight · Email agentBooking and documentation emails are drafted by the agent and sent after a single human review; every message keeps a full processing trail, so customer queries can always be answered with evidence.
Freight forwarder · Freight Agent
Electronics manufacturing · Vision inspectionThe agent orchestrates camera, lighting, edge models and inspection recipes by station; model switchover goes live the same night as the line changeover, and the miss rate is tracked as an independent quality metric, keeping inspection standards clear and verifiable.
Electronics assembly line · Industrial Visual Inspection Agent

Demo

Demo appointment

We will follow up shortly after you submit. You can also email us at service@crusbro.com.

Also available viacontact pageTelephone/email communication.

Deploy

Morphological comparison

The same set of capabilities can be split into different deployment forms based on network boundaries and compliance requirements; the following table shows common comparison dimensions in contracts and acceptance.

formconstraint
Private cloudTenant isolation, export whitelist
engine roomData does not leave the campus, offline update package
edgeIn the same segment as the camera/PLC, the time delay is written into the SLO
mixTraining at the center, inference at the edge; synchronization strategy is auditable
Deliverables
Image, model version, configuration list, rollback manual
acceptance
Freeze snapshots, differences before and after approval, session replay index
cycle
Pilot starts from 6–8 weeks (depending on data governance and system access)

White paper registration