Context + Human Authority

Reality tells AG Arch what is true.

Context explains what that reality means for the work.

Human Authority defines what the autonomous system is ultimately allowed to pursue, change, or decide.

This is where the human role enters the AG Arch model most clearly.

The purpose is not to place a person inside every execution step. It is to make sure autonomous execution remains grounded in decisions that belong to people while allowing the system to operate independently inside those decisions.

Context gives meaning to reality

A raw fact is not always enough.

Suppose AG Arch observes:

A production setting is currently `off`.

That tells us something about reality.

But it does not tell us:

• whether the setting should change;
• why it matters;
• which environment is actually in scope;
• what other systems depend on it;
• what result is wanted;
• what previous decisions affect the work.

That additional information is context.

Context can include the original intent, success criteria, scope, dependencies, applicable policies, previous decisions, constraints, and evidence requirements.

Its purpose is to help autonomous reasoning interpret reality correctly rather than treating every observed fact as an instruction to act.

Human Authority is the source of direction and permission

Agents can reason, propose plans, select tools, execute actions, and collect evidence.

But they should not become the ultimate source of their own authority.

Humans decide the higher-level questions that define the legitimate space in which autonomous work operates.

That includes decisions such as:

• what outcome is actually wanted;
• what systems or environments may be affected;
• what boundaries must not be crossed;
• what kinds of risk are acceptable;
• which decisions are too consequential or irreversible for the system to make on its own;
• when the purpose or scope of the work itself should change.

AG Arch treats those decisions as Human Authority.

The agent may determine how to work toward an authorized outcome, but it does not gain the right to redefine the outcome simply because another direction appears easier or technically possible.

Human Authority becomes a governed mandate

Human Authority does not mean that a person must remain present while the task is executed.

AG Arch turns the relevant human decisions into a governed mandate that the autonomous system can use during execution.

That mandate can express things such as:

• the scope of the work;
• allowed actions;
• prohibited actions;
• applicable constraints;
• stop conditions;
• actions that require additional authorization.

This creates a clear relationship:

Human Authority → Governed Mandate → Autonomous Execution

The human defines the legitimate boundary.

AG Arch carries that boundary with the work.

Agents and tools execute inside it.

Human Authority does not mean constant approval

A system that asks a person for permission after every routine step is not meaningfully autonomous.

AG Arch is intended to do the opposite.

Human decisions should establish enough authority in advance for normal execution to continue without unnecessary intervention.

For example, a person may authorize a task to modify configuration within one defined environment, run the required validation, and apply the approved change.

Once that boundary is established, the agent should not need another human confirmation for every file read, validation check, or ordinary execution step inside it.

Human involvement becomes necessary when the work reaches a decision that genuinely belongs outside the existing mandate.

Technical access is not Human Authority

A tool may technically allow an agent to modify several repositories, environments, databases, or services.

That access does not decide what the agent is entitled to change for a particular task.

This is why AG Arch separates:

what a capability can do

from

what Human Authority has permitted this work to do.

A more powerful model or a tool with broader credentials does not receive broader decision rights automatically.

Capability can change without changing the human-defined boundary.

Human Authority can be expressed before the task begins

The human role does not always appear as a live approval during execution.

Many decisions can already be expressed through policies, project rules, task scope, operating constraints, or previously authorized boundaries.

The autonomous system can then apply those decisions without repeatedly returning to a person.

This is important because the goal is not human supervision of every action.

The goal is human authority with autonomous execution.

The system should return to a human decision boundary only when the existing authority is genuinely insufficient — for example, when the task would need to expand into a new system, cross an established safety boundary, make an irreversible decision, or change the intended objective itself.

Context can change without changing authority

New information often appears during autonomous work.

An agent may discover a dependency, a different root cause, or an unexpected state.

That changes the context available to the system.

It does not automatically change Human Authority.

Learning that another production system is responsible for the problem does not automatically authorize the agent to modify that system.

The agent may investigate, reason about the new information, or propose a different path within its existing permissions.

But if solving the problem requires crossing the human-defined boundary, that boundary must be changed deliberately rather than silently.

Missing context and missing authority are different problems

These situations require different responses.

Missing context:

“I do not know enough to determine the correct action.”

The system needs more information or evidence.

Missing authority:

“I understand the required action, but the current mandate does not permit it.”

The system has reached a human decision boundary.

Keeping those cases separate prevents an agent from treating better information as permission, or from asking a person to approve something when the real problem is simply insufficient evidence.

Human Authority remains connected to the work as it moves

The same task may pass through multiple agents, specialists, models, or tools.

The human-defined boundary should not disappear during those handoffs.

If the work is authorized only for one environment, routing it to a more capable specialist does not expand that authorization to every environment the specialist can access.

The capability performing the work may change.

The human authority behind the work remains the same until it is deliberately changed.

Example — changing a system configuration

Continuing the same example:

The relevant context says:

• a running service is using an outdated configuration;
• the intended result is to restore the approved value;
• the service must remain healthy;
• the result must be verified afterwards.

Human Authority establishes the boundary:

• the staging environment may be changed;
• the relevant configuration may be modified;
• validation and an authorized service reload may be performed;
• unrelated systems must not be changed;
• production is outside the current mandate.

AG Arch carries those decisions into autonomous execution as the governed mandate.

Suppose the agent later discovers that the real cause of the problem is in production.

That is valuable new context.

The agent can report and reason about it, but the discovery does not give it permission to change production.

The current mandate has reached its boundary.

If production must now be changed, Human Authority must deliberately extend or replace that mandate.

How Context + Human Authority support AG Arch

Together they answer three different questions:

What is relevant to understanding this work?

Which decisions belong to the autonomous system?

and

Which decisions remain human?

Context gives the system enough meaning to reason well.

Human Authority defines the legitimate boundary of that reasoning and action.

AG Arch converts that authority into a governed mandate so routine work can remain autonomous without allowing agents to become self-authorizing.

This is the balance the architecture is designed to preserve:

human direction and decision rights, without human micromanagement of execution.

The next supporting system explains where AG Arch gets the agents, models, tools, and specialists capable of carrying out that governed work:

Capability Pool

03 — Supporting Systems