AI Agents and Increasing Machine Agency

• For most early generative AI, the user asked and the system answered.

• Agents change the verb.

• The system can act.

• That is a larger governance change than another increase in model fluency.

Action Changes the Governance Problem

Tool-using agents can already code, manage messages or calendars, search files, call services, and perform multi-step digital tasks under varying levels of autonomy.

This is current capability, not AGI speculation. Agents may read websites, emails, or documents containing instructions designed to manipulate them.

When the system only generates text, the result may be a bad answer.

When it controls tools, manipulation can become an action. This is why prompt injection becomes a security problem rather than merely a quality problem.

A system that plans, retrieves, communicates, and acts may resemble an organizational worker functionally.

Operational agency does not create ecclesial authority. The agent can act on behalf of a church without becoming the church.

An agent can be permitted to draft but not send, schedule but not cancel, propose a purchase but not execute it.

Fine-grained permissions often capture most automation benefits while preserving human control over irreversible steps.

Current, Emerging, Plausible, Speculative

Bounded organizational agents are becoming more useful for repeated workflows: metadata maintenance, internal knowledge retrieval, scheduling, public inquiry classification, routine non-sensitive processing.

The key is boundedness. Which tools? Which data? Which actions? Which escalation conditions?

Mission-specific agents could combine existing capabilities. A seeker assistant might identify language, retrieve vetted material, answer, recognize a request for human contact, and route the person appropriately.

A translation agent might prepare candidates, compare terminology, flag consistency, package review material, and escalate unresolved issues.

These workflows are technically plausible. Plausibility is not evidence of missionary fruit. Fully autonomous missionary agents, machine-run churches, or AI “finishing” the Great Commission belong in the speculative category.

The present book does not depend on them. Several agents can divide tasks: research, review, execution. This does not automatically create independent verification if all agents rely on similar models and assumptions. Diversity of system roles should not be mistaken for diversity of evidence.

Table 35.1. Agent Horizon

Horizon	Meaning	Mission posture
Current	Tool-using systems capable of bounded actions	Govern permissions now
Emerging	More integrated multi-tool organizational workflows	Pilot with tight boundaries
Plausible	Mission-specific agentic workflows with broader autonomy	Design safeguards before deployment
Speculative	Autonomous missionary or ecclesial systems	Do not build strategy on this assumption

Permission Matters More Than Prompting

As agency increases, the central governance question becomes what the system can access and do.

• Can it read email?

• Pastoral files?

• Donor records?

• Can it send, delete, publish, purchase, or modify?

• A perfect prompt does not compensate for excessive permissions.

• Give agents only the access needed for the task.

A scheduling agent does not need counseling records. A translation agent does not need financial systems. A public chatbot does not need internal personnel data.

As agents interact with services and with one another, systems need ways to establish what the agent is authorized to represent.

A church agent sending email should not appear indistinguishable from a pastor unless the role is explicit.

Identity and authorization become both technical and truthfulness issues. Untrusted content can manipulate agents, turning ordinary information ingestion into a security boundary. The long-term test is what grows around the tool. The ministry may gain speed while losing understanding, or it may use speed to create margin for better judgment and relationship. Success should therefore be evaluated over time through competence, accountability, resilience, and service rather than through the first impressive demonstration.

Memory, Auditability, and Graceful Failure

Start automation with actions that can be undone. Keep logs that allow the organization to reconstruct what happened, what tool was called, what information was accessed, and under whose permission.

Automation changes who performs the immediate action. It does not automatically change who is accountable for authorizing it.

Persistent agent memory can improve continuity and create privacy risk. Organizations should decide what is remembered, for how long, who can inspect it, and whether users can correct or delete it.

A mature agent knows when to stop. Timeouts, uncertainty thresholds, missing permissions, conflicting instructions, and high-risk content should trigger handoff rather than improvisation. The ability to refuse action is part of agent competence.