From a goal to a verified outcome.

Autonomous agents can plan, use tools, change systems, and report that a task is complete. But taking an action is not the same as producing the intended result.

In real work, several questions still have to be answered. What was the goal? What was true before the work began? What was the agent allowed to change? What actually happened after it acted? What evidence shows that the result is correct? And what should happen if reality does not match the plan?

AG Arch provides the operating and control layer around that work.

It starts with the operator’s intent and the current state of the real system. Before execution begins, it defines what success means, what authority is available, what constraints must be respected, and what proof will be required before the work can be considered complete.

AG Arch orchestrates and routes governed work across the appropriate agents, models, tools, and systems while keeping authority, evidence requirements, and the outcome standard attached to the work. Those capabilities may perform the action, but they do not decide for themselves whether the final outcome is acceptable. AG Arch separates the ability to act from the authority to define success and from the evidence required to prove it.

After an action is taken, the process continues. AG Arch observes what actually changed, collects evidence from the relevant systems, and compares the result with the success criteria defined earlier. Only when that evidence supports the intended result does the process reach a verified outcome.

If the result cannot be proven, the system does not simply treat the work as complete. It identifies what needs to change next. The plan may need to be revised, authority may need to change, a different capability may need to be used, or execution may need to stop safely.

This creates a continuous path:

Intent → Current Reality → Success Criteria → Governed Plan → Execution → Observation → Evidence → Verdict → Verified Outcome

The purpose is not to add human approval to every step. The normal path should remain autonomous inside clearly defined boundaries. Human authority is used where a decision genuinely belongs to a person, while routine execution can continue without unnecessary intervention.

AG Arch is also designed to work around existing systems rather than replace them. Business systems, infrastructure, databases, services, and other systems of record continue to hold the real state of the world. Agents and tools provide execution capability; AG Arch governs how that capability is used and how the resulting outcome is proven.

A simple example

Imagine an autonomous agent is asked to change a configuration in a live system.

Without a control layer, the agent might make the change, receive a successful response from the tool, and report that the task is done.

With AG Arch, that action is only one part of the process. The system first establishes what configuration exists now, defines what the intended result should be, determines what the agent is allowed to change, and specifies how the result must be verified. After the change, the real system is observed again. Evidence is collected, the outcome is evaluated, and only then can the work be treated as successfully completed.

That same control pattern is what the rest of this guide explains in detail.