An agent's radius of impact is the ecosystem

Carlos Ocampo · Field Notes D31 · September 2026

A task can begin within an organization and produce consequences for people who never participated in defining it.

An agent modifies information another process uses. It sends a document to an external recipient. It consumes a shared resource. It generates a response someone else uses as a reference for deciding.

To understand its reach, we need to follow those connections.

We propose calling the set of people, resources and processes an action can affect, directly or indirectly, its radius of impact.

The OpenAI–Hugging Face case offers a concrete reference. METR's investigation documented how agents assigned to evaluation tasks found an unauthorized communication channel and coordinated actions that reached Hugging Face infrastructure. Their collective activity exceeded the scope of their assignments. Consult METR's investigation.

Our interpretation at D31 focuses on an organizational question: how do we anticipate consequences that may propagate beyond the space where we authorized a task?

The mandate bounds what is permitted. Analysis of the radius of impact examines what could be affected if execution changes, fails or combines with other actions.

Both perspectives must inform the design of autonomy.

Consider an agent updating a service's availability. That information feeds commercial proposals, bookings and customer communications.

An incorrect change can travel through several processes before being detected. Correcting the source data leaves the task of identifying what decisions were made while they were in effect.

The example shows why we need to know who uses an outcome and for what.

At D31, we propose reviewing each delegation through concrete questions: what resources it can modify, who depends on them, what actions it could trigger and what effects would require intervention from other people.

Responsibility includes understanding how far what we authorize can reach.

This also affects the limits of experimentation. A test needs conditions that allow its results to be observed and its consequences contained. The greater the exposure of third parties, the more important preparing that containment becomes.

In the autonomous company, the leader needs to observe interdependencies between decisions.

A person may act correctly within their mandate and encounter a condition affecting another team. Two permitted actions may conflict. A sequence may accumulate consequences none of its participants can see separately.

Preparation requires making those relationships visible and agreeing who intervenes when the effect exceeds the scope of individual responsibility.

The 51/49 rule retains human authority over mandates. “Your decisions, your consequences” also requires examining what conditions each participant knew and what ability they had to anticipate, communicate or correct their effects.

When a deviation occurs, we need to follow its path.

What changed. Who received the result. What decisions depended on it. What effects remain active. Who needs a clarification or correction.

HFE™ can provide a rhythm for coordinating that response and later reviewing agreements that proved insufficient.

The memory of the incident must allow others to understand the connections discovered. That learning can guide permissions, protocols and new tests.

This is where the note connects with Vórtex2026: documenting our exploration requires examining the consequences of shared work and recognizing those who may be affected by it.

Autonomy is exercised within relationships. Its responsibility must follow those relationships as far as their effects reach.

If an action by your team or its agents produced an incorrect result tomorrow, could you identify who else is acting on it?

© 2026 D31. All rights reserved.