Point C1

Control Loci, Not Human Managers: An Agent-Native Routing Model

Thesis

Agent systems should route decisions to the right control locus instead of copying human management structures or escalating every uncertainty to a person.

“Authority” is a human social term. It carries assumptions about employment, accountability, hierarchy, incentives, trust, and legal responsibility. Those assumptions do not map cleanly onto agent systems. An agent with a manager title does not become wiser because of the title. A “CEO agent” is not a CEO. A “senior reviewer agent” is not senior in the human sense.

The useful question is not who has the title. The useful question is where the next decision can be resolved.

What Is A Control Locus?

A control locus is the place where a decision should be handled. In an agent-native system, common loci include:

The principle:

Delegate remediation. Escalate control-boundary changes.

Why Human Escalation Is Often Too Early

Many agent systems interrupt the human because the agent lacks a better route. It asks:

• Should I retry?
• Which source should I trust?
• Can I change this file too?
• Do you want option A or B?
• Should I continue?

Some of those questions are legitimate. Many are not.

If the task is still within scope, reversible, and evidence can be gathered, the executor or verifier should often continue. If two valid approaches exist, an arbiter can compare them. If source context is stale, a context-refresh capability can reread. If a rule is fixed, policy can decide. The human should be interrupted when the delegation boundary changes, an irreversible commitment appears, or intent/preference/accountability cannot be inferred from the delegation.

Routing Diagram

flowchart TB
    A["Pending decision"] --> B{"Inside delegation<br/>boundary?"}
    B -->|"No"| H["Human/principal or policy gate"]
    B -->|"Yes"| C{"Evidence missing?"}
    C -->|"Yes"| V["Verifier or evidence-gathering agent"]
    C -->|"No"| D{"Context stale?"}
    D -->|"Yes"| R["Context refresh"]
    D -->|"No"| E{"Multiple valid options?"}
    E -->|"Yes"| AR["Arbiter"]
    E -->|"No"| X["Executor continues"]

This diagram is intentionally simple. Real systems need domain policies, budget controls, privacy constraints, and confidence thresholds. But the routing idea is the same: do not treat the human as the default exception handler.

Examples

In coding, an agent fixing a test should run tests again after a small reversible edit. It should not ask the human whether to retry the same test. It should ask if the fix requires changing authentication or database schema when those were out of scope.

In research, an agent that finds weak evidence should search again, downgrade the claim, or surface counterarguments. It should not ask the human to read every weak source. It should ask when the thesis itself must change.

In legal review, an agent can extract clauses and classify risk under a rubric. It should not decide acceptability or send a negotiation response. The commitment boundary belongs to a human or institutional process.

Escalation-Boundary Matrix

Why This Is Not Removing Humans

Human judgment remains central, but it becomes more precise. Humans are not used as universal parsers of agent uncertainty. They handle:

• intent changes
• preference tradeoffs
• institutional accountability
• irreversible external actions
• legal or financial acceptance
• ethical or policy exceptions
• final publication, deployment, or commitment

That is a higher-value role than approving every retry.

The Risk Of Over-Automated Arbitration

An arbiter agent is not a neutral judge by default. It is another model or system with its own failure modes. It can prefer fluent explanations, miss hidden assumptions, or over-optimize for the wrong metric.

This means arbiter decisions should also be recorded. The system should preserve what alternatives were compared, what evidence was used, what criterion won, and whether the result changed the delegation boundary.

Practical Takeaway

For every interruption, ask:

1. Is this inside the existing delegation boundary?
2. Can another agent gather evidence first?
3. Can policy decide from fixed rules?
4. Is this a preference or accountability decision?
5. Would human input change the objective, boundary, or commitment?

If not, the interruption is probably a routing failure.

Claim Support

Bridge To Article 5

Control loci make long-running delegations possible. Without them, every uncertainty breaks the run and waits for the human.

Sources

• Bainbridge, “Ironies of Automation.” https://ckrybus.com/static/papers/Bainbridge_1983_Automatica.pdf
• Lee and See, “Trust in Automation.” https://journals.sagepub.com/doi/10.1518/hfes.46.1.50_30392
• Amershi et al., “Guidelines for Human-AI Interaction.” https://www.microsoft.com/en-us/research/wp-content/uploads/2019/01/Guidelines-for-Human-AI-Interaction-camera-ready.pdf
• OpenAI Agents SDK human-in-the-loop documentation. https://openai.github.io/openai-agents-python/human_in_the_loop/

Agent Involvement

This article was prepared with AI assistance from a sanitized research discussion and public sources. The human maintainer approved this publication package on 2026-06-28. Treat the design primitives as exploratory proposals, not settled standards.