FLEXVOTIXGet started
Roadmap

Measured by what must be true, not by dates.

Each phase ends when its acceptance criteria pass. These are targets for future work. None of them has been met yet.

Current status: AEGIS is pre-release, and no phase below is complete. The roadmap may change as we learn.

  1. Phase 1

    Foundation MVP

    Planned

    Scope

    • Agent registry and configuration
    • Basic multi-agent orchestration
    • Model provider integration
    • Tool execution through a gateway
    • Execution history and basic tracing
    • Basic policy and approval controls

    Acceptance criteria

    • Across 1,000 randomized crash-injected runs, zero duplicate side effects on the test tool
    • Every tool call in the test suite has a decision committed before the call
    • Property tests on 10,000 generated Mandate trees find no child outside its parent
    • Altered, expired or reused approvals are rejected
    • With the policy store unavailable, no action runs
    • Cross-tenant suite returns zero foreign rows
    • One reference plan completes on three model providers
    • Three early access teams run a non-production plan using only the docs
  2. Phase 2

    Developer platform

    Planned

    Scope

    • Visual Agent Studio
    • Python and TypeScript SDKs
    • API documentation
    • MCP integrations
    • Memory infrastructure
    • Agent testing and evaluations

    Acceptance criteria

    • A new engineer publishes a tested agent in under 15 minutes, in moderated tests
    • A user completes the quickstart unaided
    • An MCP tool schema change forces another review, and no tool bypasses Guard
    • A cross-project retrieval test returns zero results
    • A regression suite compares two versions and gates publishing
  3. Phase 3

    Production infrastructure

    Planned

    Scope

    • Bounded dynamic replanning
    • Customer-hosted runner
    • Advanced observability
    • Enterprise security work
    • Model routing
    • NVIDIA ecosystem integrations, where technically validated

    Acceptance criteria

    • Service level objective measured for 90 days before any SLA is offered
    • A network test proves that runner credentials stay in the customer environment
    • External penetration test completed, with critical findings closed
    • Metering reconciles within 0.5% of the event record
    • Each NVIDIA adapter has a reproducible benchmark before it is listed as tested
  4. Phase 4

    Enterprise scale

    Research

    Scope

    • Advanced governance
    • Private deployment
    • Enterprise identity integrations
    • Multi-region architecture
    • Advanced analytics
    • High-scale execution

    Acceptance criteria

    • A reference deployment installs reproducibly in a customer account
    • SSO and SCIM provisioning work end to end
    • Tenant data stays pinned to its region in a failover test
    • A published, reproducible load test, with thresholds set from real customer needs

Out of scope

What we will not build.

A roadmap is also a list of refusals.
  • A no-code automation product for non-engineers
  • A proprietary foundation model
  • A GPU cloud
  • A connector marketplace for its own sake
  • A consulting arm

Help decide what comes first.