Self-Hosted · Deterministic · Model-Agnostic

The enterprise runtime for governed AI agent workloads.

Oya is a deterministic execution platform for AI agents, deployed within your virtual private cloud, integrated with your approved model endpoints, and producing a complete, replayable audit record for every execution. Data remains within your security perimeter at all times.

Designed for environments governed by SEC Rule 17a-4, HIPAA, and DORA · Available directly or through systems integration partners

Talk to the founders →

execution #8412 · claims-intake-agent · replayable
001plan.compiletyped IR · 14 nodes · schema-validated
002fetch.claim_recordTRANSPARENTpolicy_id validated
003secrets.db_credentialOPAQUEisolated from model context
004llm.classify_claimendpoint: bedrock.us-east-1 (customer-managed)
005gate.human_approvalpolicy: disbursements > $10,000
006write.core_systemTRANSPARENTdeterministic ordering enforced
state integrity 100% · execution hash 0x9f2c…verified · deterministic replay
retention: immutable · policy-definedexport: SIEM / examiner-readya complete evidentiary record of agent behavior
Operates within your environment
AWS · Azure · GCPHelm & Terraform deploymentSSO / SAML · SCIM · Role-Based Access ControlCustomer-managed encryption keysPolicy-defined network egress
Platform Architecture

Correctness by construction, not by observation.

Conventional agent architectures permit the model to improvise execution, an open loop made durable by retries and recovery, and rely on observation to detect failures after the fact. A recovered loop is a resumed loop, not a correct one. Oya separates planning from execution: agent behavior compiles to a typed dataflow plan, which a deterministic runtime then executes. Non-compliant workflows are not detected at runtime. They are structurally unrepresentable.

Data Sovereignty

Deployed within your perimeter

The complete runtime operates within your cloud environment. No data path crosses your security boundary. Execution, storage, and audit records remain under your organization's control, consistent with regulatory expectations for data residency.

Model Governance

Your approved model endpoints

Route planning and execution to Amazon Bedrock, Azure OpenAI, or self-hosted inference under your own credentials. Oya does not meter, intermediate, or access your model consumption. Model selection remains a per-workload decision under your governance.

Auditability

Examination-ready execution records

Every execution produces a deterministic, replayable record with immutable retention options. Credentials and sensitive values are structurally isolated from model context. When examiners require an account of agent behavior, the evidence exists by default.

Total Cost of Ownership

The economics of internal platform development.

Where compliance requirements preclude hosted agent platforms, many organizations assemble internal platforms on open-source components: an execution engine, a framework, and internally developed governance. The engine may be free; the governance layer above it is not. The total cost of ownership of that approach warrants examination.

Internal Development
$600K–1M annually
3–5 dedicated engineers, fully loaded, ongoing
  • 75% of internally developed agentic systems are projected to fail (Forrester)
  • Open-source engines provide durability; policy enforcement, credential isolation, and audit evidence remain internal engineering obligations
  • Each model provider update introduces regression risk across internally maintained governance code
  • Typical time to first production workload exceeds 18 months
Deploying Oya
Weeks to production
annual platform license · deployed in your environment
  • Deterministic runtime available at deployment. The planning and execution architecture is the platform
  • Approximately 5× reduction in token consumption relative to ReAct-style approaches, with the method and measurements set out in the white paper
  • 100% state preservation demonstrated across six frontier models
  • Engineering capacity is directed toward agent development, not platform maintenance
Deployment Methodology

A structured path from evaluation to production.

Deployment is conducted directly with your platform organization or through your systems integration partner. Oya maintains partnerships with integrators serving financial services, insurance, and healthcare.

PHASE 1 · WEEK 0

Architecture briefing

A working session with your platform and information security leadership. Reference architecture is defined for your cloud provider, identity provider, and key management infrastructure, with security questionnaire responses addressed directly.

PHASE 2 · WEEKS 1–2

Environment deployment

Installation via Helm and Terraform modules into your cloud account. Single sign-on is integrated with your identity provider, secrets management with your KMS, and network egress policies are defined per capability. The platform initiates no external connections.

PHASE 3 · WEEKS 3–12

Structured evaluation

A 90-day paid pilot on one production-path workflow (claims intake, KYC review, or servicing triage) with success criteria defined at the outset. The pilot fee is credited in full against the first-year license upon conversion.

Architecture Comparison

Recoverability is not correctness.

Durable execution engines, including those deployable within your own environment, ensure that an interrupted agent loop resumes; they do not constrain what the agent is permitted to do within it. Agent frameworks accelerate development; they do not provide governance. Oya is the correctness and governance layer: agent behavior compiles to a typed plan before execution, and every decision is recorded as evidence.

Evaluation Criterion
Agent Frameworks & Durable Execution (incl. self-hosted)
Oya
Agent behavior model
Open-ended model loop; the agent iterates until it determines completion, with retries on failure
Behavior compiles to a typed dataflow plan before execution; non-conforming workflows fail at compilation
Scope of guarantees
Durability: interrupted workflows resume from recorded state
Durability and correctness: deterministic execution of a validated plan, with governed tool access and approval gates
Credential handling
Credentials and sensitive values pass through model context; protection is the application's responsibility
Information-flow enforcement: OPAQUE values are structurally isolated from model context
Audit capability
Workflow event histories record code execution and recorded activity results
Examination-ready records of model-driven decisions: plan, information-flow labels, approvals, and deterministic replay, with immutable retention
Governance engineering
Policy enforcement, secrets isolation, and audit tooling are built and maintained by your team on top of the engine
Delivered as platform capabilities. Governance is the architecture, not an internal project
Model economics
Token consumption governed by the iteration count of the agent loop
Customer-managed endpoints; approximately 5× lower token consumption by architecture
Engagement

Begin with a technical evaluation.

The initial engagement is a technical demonstration conducted against your own workflow: we execute it on Oya and on your current infrastructure, compare the resulting execution records, and review state integrity across both. Your architecture team defines the evaluation criteria.

Pilot Framework

90-day structured pilot

A scoped production workload deployed within your environment, with success criteria agreed in advance. The pilot fee is credited in full against the first-year platform license, ensuring executive sponsorship and defined ownership throughout the evaluation.

Discuss Pilot Scope