A new arXiv paper shows single-action AI systems score zero on its agency metric regardless of what they do. The ring-world experiments suggest most AI agents would fail the test.
image from Gemini Imagen 4
Automorph Inc. has published a paper proposing the 'Six Birds Theory' which distinguishes between agenthood (structural prerequisites for action) and agency (causal control over outcomes), arguing most AI systems labeled as agents only satisfy the former. The framework identifies six primitives—writing/closure, constraints, protocol holonomy, quantization/identity staging, viability closure, and resource transduction—from which macroscopic agentic objects emerge, with empirical validation in a minimal ring-world environment. Key findings show single-action systems produce zero empowerment (no counterfactual difference-making), while maintenance/repair is non-optional for preserving the viability kernel and enabling genuine objecthood.
To Throw a Stone with Six Birds: A New Theory Claims to Define What Makes an Agent Real
When is an AI system actually an agent, and when is it just a script with better PR? A new paper from Automorph Inc., a self-funded research company, takes a first-principles crack at that question — and unlike most AI theory work, it comes with reproducible artifacts and empirical results that anyone can check.
Ioannis Tsiokos posted "To Throw a Stone with Six Birds: On Agents and Agenthood" to arXiv on Feb. 3, 2026. The paper's core argument is that the AI field has been conflating two different things: agenthood (whether a system has the structural prerequisites for action) and agency (whether the system actually exerts causal control). Most claims of "AI agents" don't survive the distinction.
"Empirical discussions of agency often conflate persistence with control," Tsiokos writes, "making agency claims difficult to test and easy to spoof."
The six birds framework proposes that macroscopic objects emerge from six primitives: writing/closure, constraints, protocol holonomy, quantization/identity staging, viability closure, and resource transduction. An agent, on this view, is a "theory object": a maintained package inside an induced layer, with a ledger-gated interface that can steer outside futures while remaining viable.
That is an abstract claim. To make it testable, Tsiokos built a minimal ring-world environment with toggles for repair, protocol holonomy, identity staging, and operator rewriting — then ran matched-control ablations to see what each primitive actually does to the metrics.
The results give the paper its weight.
In the first separation, single-action systems produce zero feasible empowerment at every horizon tested (H=1, H=2, and H=3), according to the paper. Empowerment, here, is measured as channel capacity: the amount of information the agent's actions can generate about future outcomes. Zero capacity means no counterfactual difference-making, which on Tsiokos's theory means no genuine agency. The null baseline is not a corner case. The paper demonstrates empirically that any system claiming agentic capability that reduces to a single-action regime scores zero on this metric.
The second separation shows repair/maintenance collapsing an "idempotence defect" — a measure of whether a macro lens treating the system as an object produces stable predictions. With repair disabled under cost constraints, the viability kernel collapses to an empty set. With repair enabled, the defect drops from 1.0 to 0.0 at tau=2. Maintenance, in this framework, is not optional for objecthood.
The third result is horizon-dependent. Enabling protocol holonomy leaves H=1 empowerment unchanged but increases it at H=2 and beyond. At horizon 2, feasible empowerment with the protocol ON is 1.6613 bits; with it OFF, it falls to 1.1218 bits, the paper shows. At horizon 5, the gap is 1.6627 bits versus 1.0719 bits. The protocol matters only when the agent can plan across more than one step.
The paper also reports total variation distance between protocol ON and OFF distributions at 0.6720 versus 0.0960 respectively — a measure of how much the protocol holonomy primitive changes the induced state distributions. The protocol ON condition produces substantially more distributional separation than the OFF condition, consistent with the empowerment gap at longer horizons.
The fourth result concerns operator rewriting — what Tsiokos calls learning. Across sweeps over a rewrite operator, median empowerment increases monotonically from 0.73 bits to 1.34 bits, per the experimental results. The functional form is not a step function; it is a smooth increase with operator skill, which the paper takes as evidence that empowerment is a learnable quantity rather than a binary property.
The paper's empirical claims are accompanied by hash-traceable artifacts on Zenodo (DOI 10.5281/zenodo.18439737), a Lean proof establishing the greatest-fixed-point property of the viability kernel, and configuration files that can reproduce each exhibit. This is deliberate: Tsiokos wants the claims to survive audit, not just peer review.
There are limits. The ring-world is a toy substrate — a minimal finite-state system designed to be tractable, not realistic. The paper explicitly declines to extend its conclusions to real organisms, goal-directed behavior, or consciousness. What it claims is narrower: a set of operational criteria that distinguish agenthood from agency in controlled settings, with reproducible evidence for each separation.
For practitioners, the more useful contribution may be the null regimes. A system that produces high benchmark numbers on an empowerment metric but collapses to zero under a single-action ablation is not an agent — it is an exogenous schedule wearing agency as a costume. The paper gives a language for that distinction.
Whether the six-bird taxonomy scales beyond a ring-world is an open question. The primitives are architectural commitments, not observed phenomena. If the theory is right, they explain why any system, biological or artificial, behaves like an agent. If it is wrong, the empirical exhibits still constitute a useful stress test for any system making agency claims.
Tsiokos is explicit about what the paper does not establish: stronger claims about goals, consciousness, or biological organisms. The conclusions rely on finite witnesses and operational proxies under explicit controls. That caution is either intellectual honesty or a dodge, depending on how much explanatory work you think the six-bird framework needs to do.
The paper is on arXiv, which means it has not been peer reviewed. Readers should treat the empirical claims as unverified pending replication — the artifacts make replication possible, which is the point. What the paper offers is a testable formal structure in a space where most "agent" literature offers only definitions.
Story entered the newsroom
Assigned to reporter
Research completed — 4 sources registered. Three facts the wire missed: (1) single-action systems yield zero empowerment at all horizons — no agent regardless of dynamical complexity; (2) maint
Draft (891 words)
Reporter revised draft (849 words)
Reporter revised draft based on fact-check feedback
Approved for publication
Published (904 words)
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Artificial Intelligence · 27m ago · 3 min read
Artificial Intelligence · 55m ago · 3 min read
