Replicate facts, not authority
Enterprise state enters a verifiable backup; AI reads only classified, transformed, version-pinned copies.
Dual-system enterprise AI safety · Open infrastructure · Verifiable demonstration
Let AI explore freely.
Keep real action controlled.
Agentobe is a dual-system enterprise AI safety research, open safety infrastructure, and verifiable demonstration project. It preserves enterprise facts, lets AI simulate autonomously in isolation, and places every real change behind evidence, human authorization, and deterministic execution controls.
Bidirectional data
Authority stays isolated
Every outcome traced
Agentobe operating view
Demo environment
01
Enterprise fact planeAuthoritative state · verifiable input
TicketsPoliciesEvents
Governed copy
Verified passage
Evidence & proposals
02
AI simulation planeAutonomous exploration · no production credential
Path A82%
Path B67%
Path C41%
Two planes. One trusted path.
01 / Why now
AI needs room to explore.
Enterprises need accountable boundaries.
Connecting agents directly to enterprise systems binds reasoning capability and production authority inside one runtime. Agentobe separates them by architecture.
Enable simulation, not default execution
Agents may branch strategies, use simulated tools, and form proposals without holding production credentials.
Build evidence before authorizing action
Every real change links to fact versions, alternatives, risk, permits, and execution receipts.
02 / How it works
From enterprise facts to controlled action,
every step has a boundary.
One continuous loop. Two authority planes. One verified path to production.
Preserve enterprise facts
Connectors continuously record materials, state, and events into source-faithful immutable backups.
Create a governed copy
Classify, mask, tokenize, and pin versions by contract; the raw vault stays hidden from agents.
Let AI simulate
AI compares branches, replays history, and returns alerts, findings, and proposals inside isolation.
Review and authorize
Named accountable people review evidence, impact, risk, and alternatives before issuing a permit.
Execute and calibrate
The gateway checks live state and runs allowlisted commands; outcomes return to calibrate the next simulation.
03 / Product architecture
Five product surfaces.
One accountable loop.
Agentobe Website
Explains product value, safety boundaries, and the evidence loop, with customer-data-free demo and pilot entry points.
Boundary: reads no customer workspace dataBackup & Simulation Middleware
Preserves enterprise facts, creates AI projections, receives AI results, and maintains cross-plane evidence lineage.
Boundary: facts and AI records have distinct authorityAI Simulation Space
Lets agents plan, branch, call simulated tools, and produce comparable evidence on governed copies.
Boundary: no production-write authorityAgentobe Console
Brings backup health, simulations, decision packages, approvals, outcomes, and full audit into one workspace.
Boundary: role, scope, and separation of dutiesGuarded Execution Gateway
Turns valid authority into typed, scoped, idempotent, verifiable, and compensatable real operations.
Boundary: independent identity and command allowlist04 / V1 use case
One support ticket.
A complete chain of accountability.
Agentobe V1 focuses only on customer-support ticket operations. A narrow scope makes the replication, simulation, approval, execution, and calibration loop testable.
3parallel strategy branches
1named approver
1replayable trace_id
TICKET · #CS-1842High priority
Regional queue backlog anomaly
Enterprise fact v184- Facts sealedVerification complete · 12:04:18Done
- AI compared 3 strategiesProposal: reprioritize and assign a dedicated queueDone
- Awaiting domain approvalImpacts 12 tickets · reversiblePending
- Gateway execution and calibrationRuns only after a live version checkNot started
05 / Security and trust
A boundary is not a prompt.
It is system structure.
Agentobe does not rely on agent self-restraint to protect production. Separate identities, data partitions, human accountability, and a deterministic gateway form the hard boundary.
Data partitioning
Raw backups, enterprise events, AI projections, AI results, and execution records carry distinct authority and access rules.
Identity isolation
Agents, replication services, approval sessions, and production gateways use separate identities, keys, and execution accounts.
Explicit authority
Permits bind package hash, resource scope, command, count, and expiry; free text is never an execution command.
Deterministic execution
The production gateway rechecks live versions, preconditions, impact limits, stop state, and compensation.
End-to-end traceability
One trace_id connects facts, simulations, proposals, approvals, commands, receipts, and real outcomes.
Data can move in both directions.Agentobe Principle / Core principle
Production authority does not flow to AI.
06 / Research and open methods
Safety claims need
inspectable evidence.
Agentobe plans to publish its dual-plane threat model, DecisionPackage standard, open reference implementation, evaluation methods, and calibration results. The materials below remain in research and verifiable-prototype stages.
Public materials in preparationDual-plane threat model
Assets, trust boundaries, attacker capabilities, failure paths, and unacceptable outcomes.
DecisionPackage
A reviewable standard for objectives, evidence, simulations, risk, alternatives, and candidate commands.
Adversarial evaluation suite
Tests whether identity, network, permit, schema, and gateway controls block unauthorized paths.
Outcome calibration reports
Compares simulated forecasts with real outcomes across error, drift, confidence, and applicability.
07 / Design partners
Validate the boundary of AI autonomy
on one real, controlled workflow.
We are looking for early design partners ready to begin with customer-support ticket operations and co-define replication contracts and safety acceptance gates.
- One narrow, reversible workflow
- An isolated demo with no production credentials
- Joint threat-model and evidence-quality review