An operating foundation for governed AI work.

Define the objective and the authority around it.

The control plane is intended to give teams a place to configure workflows, identify accountable owners, manage identities and permissions, define budgets and inspect execution.

Planned responsibilities include:

• Versioning workflow definitions and approval rules.
• Determining which data, tools and actions a workflow may use.
• Setting resource limits and conditions for escalation.
• Reviewing evaluation evidence before approving changes.
• Connecting execution records with owners and business objectives.

Coordinate the work beneath the workflow.

Agents OS is our name for the planned distributed orchestration and resource-management layer. It is not a replacement for a general-purpose operating-system kernel.

The intended execution model would:

Parallel work is useful only when it preserves correctness and the intended business result.

• Break work into tasks with explicit dependencies and completion conditions.
• Assign appropriate agents, models, tools and workers within defined constraints.
• Run independent tasks in parallel without ignoring ordering or shared-state requirements.
• Preserve progress across handoffs and recover from failures.
• Apply safeguards against repeated external actions during retries.
• Manage concurrency and backpressure rather than overwhelming connected systems.
• Pause for human approval when an action exceeds delegated authority.

Operate the foundations without separating them from the work.

The intended infrastructure layer covers agent runtimes, model serving, enterprise context, connectors and compute operations. Capacity management, maintenance, upgrades and recovery are part of the operational vision.

Specific hosting locations, deployment models, service levels and integration availability are not confirmed. Discuss requirements with us rather than assuming a particular deployment is supported.

Agent lifecycle

Planned support for creating, integrating, versioning and operating business agents as part of governed workflows.

Models suited to the task

We intend to explore small language models (SLMs), adaptation and model selection where quality, latency and cost make them appropriate. Smaller does not automatically mean better, cheaper or sufficiently accurate.

Permissioned enterprise context

Retrieval-augmented generation (RAG) is intended to connect answers to authorised business information. Access rules and source quality would still determine what the workflow may use.

Tools and connectors

Our connector direction includes the Model Context Protocol (MCP) as one way to expose tools and data. No named business-system integration is claimed available.

Evaluation and observability

Planned evidence would connect workflow versions, actions, approvals, quality checks and resource usage. Evaluation should identify regressions, not simply count activity.

Identity, tenancy and cost controls

Tenant identity, data permissions, execution state and resource limits are intended to be governed together. Implemented isolation, billing packages and budget-enforcement behaviour remain to be validated.

Our proposed loop is to define a result, execute within bounds, collect permitted evidence, evaluate performance, propose changes, and test before release.

Customer policy should determine approval and rollback requirements. Learning from execution does not imply unrestricted collection of customer information, reuse across tenants or permission for an agent to rewrite its own production authority.

Start with the business objective, the systems involved and the decisions that need human approval. We can discuss the platform direction and the questions that would need to be resolved.