Platform

The infrastructure layer for agentic work.

Minato connects the operating surface to durable workflows, specialist agents, policy gates, evidence, and Kubernetes ownership boundaries.

Ownership topology

Four layers, with a visible owner at every boundary.

Product intent moves into the control plane and tenant runtime. Status, approvals, typed outputs, artifacts, and run evidence move back toward the operating surface.

Minato / runtime topologyDesired state ↕ observed evidence
  1. 01
    Operating surface

    Product surfaces

    The place where operators direct work, make decisions, and inspect what the system produced.

    • Director
    • Approvals
    • Workflows
    • Results
  2. 02
    Minato control plane

    Minato control plane

    The control boundary that translates authored intent into reconciled workspace and workflow state.

    • Workspace lifecycle
    • Orchestration
    • Policy gates
    • Model access policy
  3. 03
    Tenant namespace

    Tenant runtime

    Namespace-scoped runtime components execute work and retain the state and artifacts that belong to the workspace.

    • Workspace control
    • Specialist agents
    • Model gateway
    • Services
    • Store + artifacts
  4. 04
    Kubernetes foundation

    Kubernetes foundation

    The operator reconciles declared resources into the cluster boundary selected during deployment planning.

    • Helm
    • Operator
    • Namespaces
    • Secrets + storage
Reconciliation: the operator turns declared workspace state into namespace-scoped resources.Evidence: workspace-control records lifecycle events and projects the current run and artifact views.
Durable execution

Workflows keep structure around long-running work.

A workflow is an authored task DAG: dependencies establish order, gates capture decisions, and the ledger preserves the evidence produced as work advances.

Workflow / runDeclared path
  1. 01 / DefineTyped inputs

    Bind the objective and the outputs each step may expose.

  2. 02 / BranchSpecialist work
    ResearchDraft
  3. 03 / GateHuman decision

    Resolve a structured approval before dependent work.

  4. 04 / JoinResult + evidence

    Converge outputs and retain the resulting artifacts.

Branches and joins
Dependencies let a workflow separate independent specialist work and converge it into downstream steps.
Retries and schedules
Schedules create new workflow instances, while retry policies can target known error classes with bounded backoff.
Typed outputs
Steps declare parameter and artifact outputs so downstream references can be validated before a run starts.
Human gates
Approval steps can pause execution for a structured decision before dependent work proceeds.
Append-only run evidence
Lifecycle events are added to the run ledger, then projected into current run, question, and artifact views.
Model gateway

One explicit provider boundary for agent runtimes.

The namespace-local gateway centralizes provider credentials and exposes configured endpoints to agent runtimes. Provider connections are selected deliberately for each workspace.

Current gateway providers are Anthropic, OpenAI, Gemini, and Codex.

  • 01Anthropic
  • 02OpenAI
  • 03Gemini
  • 04Codex
Deployment boundary

Plan the environment before the operator reconciles Minato into it.

Minato is Helm-packaged and operator-managed for Kubernetes. During deployment planning, we validate the target distribution, storage, ingress, identity, and policy controls before committing to the environment.

Deployment planning inputs
Distribution
Kubernetes version, distribution behavior, and upgrade path.
Storage
StorageClass behavior, persistence needs, and recovery plan.
Ingress
Ingress controller, DNS, TLS, and exposed service boundaries.
Identity
Human and service identities at the API and runtime edges.
Policy
Namespace, secret, network, and workload policy controls.
Security posture

The current install is intended for trusted or privately controlled environments. Internet-facing access requires the authentication, ingress, identity, and policy boundary to be designed explicitly for that environment.

Let's talk

Start with one operating problem.

Book a call to identify the workflow, the human decisions around it, and the Kubernetes boundary it should run inside.

Book a call