Skip to content

Our vision

Agents that improve.
Control that stays yours.

We believe the future of AI should be shaped by the people who use it. Individuals and organizations should be able to run, adapt and share agents without depending on one model or harness provider.

Orchestrator Zero is our open foundation for that future, from one person’s machine to a company’s infrastructure.

See what openness means
The model supplies the reasoning.
Hosted modelsFrom model providers
Self-hosted modelsOpen models you serve
Compatible model services · hosted outside oz0
Orchestrator Zero

Runs agents. Applies your rules.

Agent runtimePlugin tools + skills

oz0 runs the model → tool → result loop.
External harnesses are optional.

Work runs on machines you control
Your applications, infrastructure and data

You operate the infrastructure, credentials and integrations. Model training and serving stay outside oz0.

Openness, in practice

Build once.
Keep your choices.

People and organizations should decide how their agents run, what they may do and who they rely on. A shared foundation gives small teams and large companies a way to build on each other’s work.

Choose the model
Use an external provider or serve an open model yourself. Switching to a compatible model should preserve your agents, tools and rules. Test quality before switching.
Run the agents yourself
oz0’s default runtime manages context, calls tools and continues the job across compatible models. Other runtimes, such as OpenCode or the Claude Agent SDK, are optional.
Build on shared work
Our vision for oz0 Hub is a place to discover, improve and redistribute plugins. Share improvements under their licenses, or keep your custom work private. Plugins can also come from repositories or local files.

The direction

From useful jobs to improving systems.

These are stages of our vision, with no promised release dates. Each step depends on the reliability of the one before it.

Explore the product today
  1. TodayImplemented foundations

    Run work with boundaries.

    The default runtime runs agents across compatible model providers. Jobs, rules, approvals, budgets and history keep the work accountable. Supported comparisons and reviewed plugin-change proposals provide a starting point for improvement.

    Public software access is coming soon. The plugin improver’s coding sessions are currently tested against a scripted model. See testing limits

  2. NextDirection

    Grow a shared library of useful work.

    Plugins can package agents, skills and tools. We want a growing ecosystem around oz0 Hub where people publish useful work, adapt it to their needs and contribute tested improvements.

    Agent
    Takes on a task
    Skill
    Provides procedures and examples
    Tool
    Performs an action

    The ambition is a thriving ecosystem. Public software and Hub access are coming soon.

  3. As evidence growsDirection

    Give proven work a carefully defined scope.

    Owners use successful runs and failure tests to decide which specific actions may run without asking. Agents can propose better rules; an authorized person approves any change in authority.

    • AllowRead approved logs
    • AskChange production
    • RefuseErase audit evidence

    Trust applies to a task and scope. It can be reduced after a failure. Records and spending limits stay in place.

  4. Longer termAmbition

    Agents help build what comes next.

    We want agents to create plugins, improve existing ones and operate recurring work. People define outcomes and boundaries, review the evidence and choose which improvements to share. Those improvements become the starting point for someone else.

    Work continues between human reviews, within assigned budgets and stop conditions. Improvement means changing instructions, tools and software; model training stays with the model developer.

The loop we’re building toward

Every new version has to earn its place.

For example: an agent proposes fewer redundant checks in a network plugin. The new version must still find the same faults and respect the same access rules.

  1. 1

    Observe

    Find repeated failures or wasted work.

  2. 2

    Propose

    Agent drafts a new plugin version.

  3. 3

    Evaluate

    Test outcomes, security and cost in isolation.

  4. 4

    Authorize

    An owner accepts or rejects the change.

  5. 5

    Release

    Roll out, monitor and keep a rollback path.

Results feed the next proposal. Rejected changes return for revision; previous evidence remains traceable.

Keep the evidence.

Preserve the link between a job, its plugin version, actions, approvals and outcome under an explicit retention policy. More autonomy must not erase accountability.

Optimize for useful work per euro.

Compare models and instructions on the same tasks. Use smaller models where they pass, reduce wasted context and retries, and stop unproductive loops. Count infrastructure and review effort alongside model spend.

Why we believe this direction is worth building.

Research informs the vision. It does not establish that all coding will become autonomous, or that oz0 has achieved these future capabilities.

OpenAI · February 2026Harness engineering
An engineering team describes agents writing its software while people shape the environment and feedback loops.
Anthropic · October 2025Equipping agents with Skills
Reusable instructions and resources let agents draw on specialized procedures.
Anthropic · January 2026Evaluating AI agents
Evaluate the model and harness together, and check the actual outcome beyond the agent’s own report.
Google DeepMind · May 2026AlphaEvolve’s growing applications
Evaluated code improvements have reached real infrastructure and research. This supports a direction, not unrestricted self-improvement.

We also draw on Anthropic’s work on separating execution from model services and keeping people in control. Sources reviewed 4 November 2026.

An Apache 2.0 core, with public access coming soon.
Built for people to run, adapt and build on.