← Back to Home

What separates persistence from control

To Throw a Stone with Six Birds: On Agents and Agenthood

Separates being a stable object (agenthood) from making a measurable difference (agency), and tests both in a controlled toy world.

Preprint - v1Not peer reviewedPublished Jan 31, 2026Open access

Plain-language overview

The paper draws a sharp line between agenthood (persisting as a stable, maintained object) and agency (being able to change outside outcomes by choosing different actions). It formalizes agents as packaged objects with budget-gated interfaces, then tests the framework in a minimal ring-world where every quantity can be computed exactly. Side-by-side comparisons show how removing choice, adding repair, changing operation order, and learning new skills each affect survival and causal impact.

Agency is difference-making under constraints, not a story inside the agent.

- Ioannis Tsiokos

At a glance

Agenthood vs agency

Persisting as a stable object (agenthood) is a separate question from being able to change outcomes (agency). You can have one without the other.

Viability and packaging

Viability: the largest set of states from which the agent can always survive. Packaging stability: does the agent hold together when re-described through a coarser lens?

Feasible empowerment

Measures how much the agent's choices can change what happens next, counting only actions the agent can actually afford.

Nulls and controls

Every claim is compared against a no-choice baseline. If the baseline also shows the signal, the claim is an artifact.

Core lens

How the paper is structured

A small set of primitives, each with explicit audits and controls.

What exists

Theory layer

A descriptive layer defines which actions and outcomes are meaningful at the coarse scale. Without a layer, "action" has no meaning.

What persists

Agent as theory object

A persistent, maintained package with a budget-gated boundary. It survives re-description and has guaranteed survival states.

What matters

Difference-making

Feasible empowerment measures how much choosing different actions changes what happens outside the agent.

Highlighted results

What the paper establishes

Each claim is paired with a control or audit.

No choice means zero agency

When only one action is available, or when a hidden schedule controls everything, causal impact is exactly zero.

Repair makes the agent a stable object

Switching on self-repair drops the packaging instability score from maximum to zero: the agent becomes a coherent, re-describable "thing."

Operation order extends long-range control

Varying the order of operations increases causal impact at longer time horizons without affecting single-step control.

Learning increases causal impact

Letting the agent rewrite its own transition rules raises its causal impact monotonically as skill improves.

Methods and reproducibility

How the results are supported

  • Exact finite computations in a ring-world where every quantity can be verified.
  • Survival sets (viability kernels) computed as greatest fixed points over reachable successors.
  • Packaging stability measured by applying the coarse lens twice and checking for drift.
  • Side-by-side comparisons with configuration hashes and audited output files.

Sanity checks

  • - No-choice regimes keep causal impact at exactly zero.
  • - Schedule traps appear only when the model mistakenly treats the schedule as a controllable action.
  • - Packaging instability drops to zero only when self-repair is switched on.

Media-ready

Artifacts and audit tables

The repository regenerates every figure and table, with hashed configurations and metric logs for each run.

- Causal impact sweeps and no-choice controls

- Survival set computation artifacts

- Packaging stability measurements across repair settings

Regenerate figures from code

Limitations and scope

Read-this-first caveats

  • Status: research preprint, not peer reviewed.
  • Finite proxy metrics in a toy world, not direct claims about real-world agents.
  • The toy substrate isolates mechanisms; it is not a model of biological organisms.
  • No claims about goals, consciousness, or open-ended evolution.

Citation

How to cite

Ioannis Tsiokos (2026). To Throw a Stone with Six Birds: On Agents and Agenthood. Zenodo. https://doi.org/10.5281/zenodo.18439737

BibTeX

@misc{tsiokos2026throw,
  title = {To Throw a Stone with Six Birds: On Agents and Agenthood},
  author = {Tsiokos, Ioannis},
  year = {2026},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.18439737},
  url = {https://doi.org/10.5281/zenodo.18439737}
}

Press and contact

Talk to the author

For media inquiries, figures, or walkthroughs of the artifacts, reach out directly.

Ioannis Tsiokos

ioannis@automorph.io

Corresponding author - Press contact

Questions welcome about the agenthood/agency distinction, causal impact measurement, and no-choice controls.