FLEXVOTIX
Sign inGet started
Get started

AEGIS / Build. Orchestrate. Execute. Govern.

The infrastructure for autonomous AI.

Build, orchestrate and operate intelligent agent systems with the infrastructure to plan, collaborate, execute and evolve across enterprise environments.

Explore AEGISView platform architecture

In development FLEXVOTIX is building AEGIS. It is pre-release.

Governance / AEGIS GuardRoot MandatePlanDelegatedStepExact approved actionObjectiveApps, SDK, consoleOrchestratorPlan v1 / 6 stepsResearchAnalysisVerificationDecisionExecutionapprovalOne gatewayERP ledgerVendor masterTicketingMCP serversData warehouseTools and enterprise systemsModel and inferenceReasoningFastOpen modelsEnterprise-hostedObservability / AEGIS TraceObjectiveOrchestrator: Plan v1, 6 stepsAgentsResearchAnalysisVerificationDecisionExecutionOne gatewayTools and systemsERP ledgerVendor masterMCP serversData warehouseModel and inferenceReasoningOpen modelsObservability

Select a layer to isolate it. This map illustrates the designed architecture; it is not a live system.

The platform

From individual agents to coordinated intelligence.

A model can answer, and a prompt can instruct. Neither one can resume a plan after a crash, limit what a sub-agent may do, or show what happened last Tuesday. Production agent systems need infrastructure. AEGIS is designed to connect it in one platform.
Execution Record: every capability reads from and writes to it

With models and prompts alone

A prompt can describe a plan, but nothing validates it or resumes it after a crash.

What AEGIS is designed to add

Objectives become versioned plans, checked before any step runs.

Read about planning

Signature demonstration

See autonomous intelligence in motion.

A simulated finance team asks AEGIS to review its vendor invoices. Start the run, watch authority narrow, and decide two payments yourself.
Simulated. Synthetic data. No real systems connected.

Autonomous Enterprise Operations

Objective Analyze outstanding vendor invoices, identify overdue payments, verify vendor information, prepare payment recommendations and request approval.

  1. Objective receivednot started
  2. Planner decomposes the tasknot started
  3. Research retrieves synthetic recordsnot started
  4. Analysis evaluates invoicesnot started
  5. Verification checks recordsnot started
  6. Decision prepares recommendationsnot started
  7. AEGIS Guard evaluates permissionsnot started
  8. Human approval requestednot started
  9. Authorized execution simulatednot started
  10. AEGIS Trace records the historynot started
AuthorityPlannerIdleResearchIdleAnalysisIdleVerificationIdleDecisionIdleGuardIdleHumanIdleExecutionIdleTraceIdle0 s5 s10 s15 s// 252.0 s wait272 s
Authority heldWork doneWaiting, no compute heldStopped or denied
PlannerIdleResearchIdleAnalysisIdleVerificationIdleDecisionIdleGuardIdleHumanIdleExecutionIdleTraceIdle

One objective, eight synthetic invoices, $51,300.00 outstanding.

Watch the plan unfold, then decide two payments yourself. Your choices change the outcome. Nothing here connects to a bank, vendor system or database.

Activity timeline / exe_demo_A7K2

Events appear here as the simulated run progresses.

Multi-agent orchestration

One objective. Multiple intelligent agents.

AEGIS is designed to compile an objective into a validated plan, then coordinate agents through it. Select a concept to see where it lives in the invoice plan.
ObjectiveRoot MandatePlannerCompiles planResearchInvoicesResearchVendorsAnalysisFinds overdueVerificationvendor.verifyVerificationpo.matchDecisionProposes onlyGuardDecidesHumanApprovesExecutionExact actionsObjectiveRoot MandatePlannerCompiles planResearchInvoicesResearchVendorsAnalysisFinds overdueVerificationvendor.verifyVerificationpo.matchDecisionProposes onlyGuardDecidesHumanApprovesExecutionExact actions

Task decomposition

The Planner turns one objective into a versioned plan of steps. A plan is data, so it can be validated before any step runs.

How orchestration works

Unified platform

Eight capabilities. One execution record.

AEGIS is designed as one platform, not eight products. Every capability reads from and writes to the same record, so state, trace and audit do not drift apart.

Build

Execute

Act and think

Control and prove

How AEGIS works

The life of an objective.

Eight stages, from the request to the evidence. Each one writes to the record, so nothing has to be reconstructed later.

Business Objective

A person or application submits an objective, with stated limits.

Written to the record
objective, principal, root Mandate

Developer experience

Designed for builders of autonomous systems.

Authority is an API object, not a settings page. Request a Mandate, submit an objective, and handle a denial as a typed result.

  • Agent APIs

    Versioned agents, tasks, executions and approvals, in one REST design.

    Preview
  • Developer SDK

    Python and TypeScript, generated from the API definition.

    Coming soon
  • Tool integrations

    HTTP tools with typed schemas and declared side effects.

    Planned
  • MCP connectivity

    An MCP client, with every tool still passing through Guard.

    Planned
  • Model provider compatibility

    Native and OpenAI-compatible adapters, using your own keys.

    Planned
  • Documentation

    Concepts, data model, error codes and API design are readable now.

    Documentation in development
  • Sandbox

    Experiment with synthetic tools and no credentials.

    Coming soon
Read the design docsDeveloper overview
API design preview. Not yet installable.
# API design preview. This package is not published.
from aegis import Aegis, PolicyDenied

client = Aegis()

run = client.tasks.create(
    agent="invoice-review@4",
    objective="Review overdue vendor invoices",
    mandate={
        "tools": ["records.search", "ap.schedule_payment"],
        "unattended_limit_usd": 5000,
        "expires_in": "24h",
    },
)

try:
    result = run.wait()
except PolicyDenied as denied:
    # A denial is a typed result you can handle in code.
    print(denied.reason)  # vendor_not_in_master

Enterprise applications

Built for work where the action has consequences.

Agentic workflows earn their place where a task spans systems and some steps need a person. These are example patterns that AEGIS is designed to support. None is a customer deployment.
Example pattern, not a customer deployment

Reconcile vendor invoices and prepare payments.

Runs unattended, within limits

Retrieve records, match POs, flag anomalies, schedule low-value payments to verified vendors.

Requires a person

Payments above a limit, and any change to bank details.

Intended value: Finance staff review exceptions, with evidence attached, rather than every line.

Model and infrastructure ecosystem

An open architecture for evolving AI.

Agents reference routes, not vendors. AEGIS is designed to support multiple model and infrastructure ecosystems, and we label each by what has actually been shown.
TechnologyWhere it would fitCategoryTargetStatus today
OpenAI-compatible endpointsOne adapter for any endpoint that speaks this API shape.InferencePhase 1Planned
Direct model provider APIsNative adapters for commercial model providers, using your own keys.ModelsPhase 1Planned
Open-source modelsReached through any compatible self-hosted or managed endpoint.ModelsPhase 1Planned
Enterprise-hosted inferenceModels behind your own network boundary, reached through a private connector.InferencePhase 4Research
NVIDIA NIMContainerized inference microservices, reached through the OpenAI-compatible adapter. Production use needs the customer's own NVIDIA AI Enterprise license.InferencePhase 3Evaluating
NVIDIA NemotronOpen models as a selectable route for planning and tool-calling steps, once benchmarked.ModelsPhase 3Evaluating
NVIDIA NeMo GuardrailsAn optional content and tool-call check inside a Guard decision. It complements Mandates and policy.SafetyPhase 3Evaluating
NVIDIA NeMo Relay and HelixCandidates for agent instrumentation and for evaluation over exported records.Evaluation and telemetryPhase 3Evaluating
NVIDIA NeMo FabricA candidate route for running external agent harnesses under the gateway.Evaluation and telemetryPhase 3Research
NVIDIA Dynamo-TritonFormerly Triton Inference Server. For serving embedding and classification models, if a workload justifies it.InferencePhase 3–4Research
OpenTelemetryTrace export through an adapter. The GenAI conventions are still in Development status.Evaluation and telemetryPhase 2Planned

No integration listed here exists yet. All are planned, evaluated or under research. NVIDIA, NIM, Nemotron, NeMo and Dynamo-Triton are trademarks of NVIDIA Corporation. They are named here to describe planned technical compatibility. FLEXVOTIX has no partnership, certification or endorsement from NVIDIA.

All planned integrations and the evidence ladder

Governance and observability

Give agents the work. Keep the authority.

Autonomy is only useful inside defined boundaries. AEGIS is designed to keep those boundaries outside the model, and to record every time they were tested.

AEGIS GuardAEGIS Trace

Sample data from the simulated invoice run.

Authority narrows at each delegation. Band width shows the scopes held.

  • Root Mandate records, vendors, payments ≤ $5,000 unattended
  • Execution agent ap.schedule_payment, released actions only
  • Verification agent vendor.verify, po.match, read only
  • Research agent records.search, read only
  • Decision agent No tools. May only propose.

Technical architecture

What runs where, and what happens when it fails.

Execution runs from top to bottom. Memory, governance, security, observability and evaluation surround it. Select any part.
Architecture target

Orchestration Engine

Keeps plans as data, resolves dependencies, schedules steps and derives each child Mandate.

When it fails
If a step fails, its retry class and failure policy decide what happens next.

Build what autonomous AI makes possible.

Explore the infrastructure for building and operating intelligent multi-agent systems.

Explore AEGISGet in touch