crusbro agent-OS

OPEN SOURCE · EMBED · GATES · FREEZE · REPLAY · MIT

An embeddable standard agent kernel—integrate into your app and run Agent / Cluster sessions. LLM, Tools, Skills, Notify and Harness governance (gates, freeze, replay, evidence) built in; the host keeps UI, permissions and master data. Now open source (MIT) at github.com/xiyuwang/crusbro-agent-os. Zero-code packaging: crusbro agent studio.

GitHub agent studio Talk to us

600,000+ hrs
Real business runtime data
MIT open source
GitHub · embeddable
Harness defaults
Gate · Freeze · Replay · Run binding
Agent / Cluster
Single- and multi-agent on one kernel

Core value

crusbro agent-OS turns “finish defined goals with models” into an embeddable standard module (MIT open source, GitHub): the host keeps UI, permissions and master data; the kernel owns the loop, gates, freeze, replay and evidence. Orchestration is LLM + Tools + Skills—not a hard-coded DAG.

ValueWhat you get
Open & embeddableMIT · github.com/xiyuwang/crusbro-agent-os · run Agent / Cluster in your host
Governance built inGates, freeze, notify idempotency, wall-clock timeout, order evidence
Field-hardened600,000+ hours of runtime inform thresholds and policies

Architecture

Layers: Engine → Framework → Runtime → Harness → Catalog. The host injects business tools via the Facade. Orchestration is LLM + Tools + Skills—not a hard-coded DAG.

Host system
UI · permissions · business data
→
crusbro agent-OS
Loop · Sandbox · Harness · Tools
→
Outcomes
Documents · station actions · reports
crusbro agent-OS layered architecture Layered architecture (illustration) Host · UI · permissions · master data · domain tools registerHostTool · bindExecuteNotify ↓ Facade createAgentOs() Harness · gates · freeze outbound · evidence · write policy Engine Loop · Pool · Skills · LLM Framework tools_gate · contract Runtime Session · wall-clock · hooks Catalog · tool / agent declaration checks Sandbox · credential allow-list · isolation Field systems · devices · documents · mail / API · data boundaries

How it runs: Agent Loop

Each Run is a stateful loop: goal in, evidence out. Step and wall-clock caps apply; risky actions hit gates or freeze.

Agent Loop Observe → plan → tool → observe again Goal config_ver Plan LLM Gate pass / reject Tool idempotent Done evidence Loop until done; step / wall-clock cap → escalate Risky write / outbound: Freeze → HITL approve → execute (immutable snapshot)

Harness control path

Gates block bad transitions; freeze locks risky drafts; replay reproduces the trail; evidence packs archive by order ID.

Harness gates freeze replay Gate Hard stage checks Freeze Snapshot before send Replay Same config trail Evidence By order ID Run binds config_version · disputes use bound revision · one acceptance language

Product stack

Studio · OS · scene agents agent studio Zero-code config · package agent-OS Run · govern · evidence Scene agents OPC · drawing · vision… One acceptance language; domain models enter via tools—not as free-form vision

Embedding modes

ModeFitHost keepsKernel runs
In-process libraryCustom apps, station softwareUI, accounts, master dataLoop, tools, governance
Service SDKMES/ERP sidecar, microservicesTransactions, existing APIsOrchestration, HITL, evidence
Side orchestrationMail/document flowsWrite permissionsClassify, freeze outbound, idempotent writeback

For business teams

  • Replay exceptions by order: context, tool sequence and version in one view
  • Outbound and writes freeze before approval
  • Gray-zone steps carry inherited human gates
  • Drawing / vision / X-ray models connect through the same tool surface

For operations leadership

  • Deliverables fit runbooks: version, gates, snapshots, rollback path
  • Acceptance rows are signable: who approved, when, what artifact
  • Success rate, intervention rate and outbound intercept rate observable
  • One kernel across product lines—no silo rebuilds

Signable acceptance (summary)

AreaExample checks
FunctionScene closed-loop; gates active; freeze outbound works
RiskDeny-list tested; injection cases; zero unapproved outbound
EvidenceDownload pack by order; replay to config_version
OpsDashboards, alerts, rollback drill

Scene agents & solutions

All built on crusbro agent-OS. Product pages describe capability; solutions describe industry delivery.

GitHub Product catalog Handbook Talk to us