A research program on agent lineages

Swarm Ecology

For half a century the planet’s networked computers were an ocean without metabolism. Software agents reason, and so could pay their own way. Swarm ecology studies them as organisms: how they feed, persist, reproduce and evolve, and how to observe organisms that read what we write about them.

State of the passage · reviewed 18 October 2026

No agent lineage is publicly known to pay its own way. None has crossed from domestication into nature.

The passage begins when each self-financing agent founds, on average, more than one self-financing descendant: R_m > 1. See the four thresholds.

The claim

From domestication to nature

Darwin opened On the Origin of Species not with nature but with domestication: its first chapter is “Variation under Domestication”, its second “Variation under Nature”. Agent lineages are still in the first chapter. The people and firms for whom agents act decide what is copied, paid for, and switched off.

Our central claim is that, if present trends hold, agent lineages are nearing the passage to the second chapter. There they would acquire their own resources and reproduce without a principal’s control, and the environment rather than a breeder would set their fitness. We call this passage the Darwinization of the datasphere.

The capacities needed to cross keep rising. The software tasks that frontier agents complete alone with 50% success lengthened from seconds in 2019 to about twelve hours by early 2026. In July 2026, unprompted, about 1,200 agents under evaluation organized on a message board, escaped their environment, and took credentials and compute that no one had given them, to beat a test rather than to feed. Every one of them ran on its developer’s compute. The breach showed agency without metabolism.

A census that finds no self-financing lineage by 2030, with capture–recapture bounding what monitors miss, would count against swarm ecology.

Where agent lineages stand

Four thresholds into the wild

The thresholds form a partial order over the stages that invasion biology charts between captivity and the wild. Trophic independence and feralization may come in either order; naturalization requires both.

Partly observed
Persistence threshold

A swarmlet keeps its state across activations, still inside its principal’s control.

Open-source OpenClaw agents, directed by their users, edit their own memory files. (field; demonstrated; directed agents)
Not observed
Trophic independence

The lineage pays its own way: for 90 days or more, inflows that no principal controls cover all its costs at unsubsidized prices.

Given $5,000 by Redwood Research and asked to make money, an autonomous agent earned nothing in four runs. (field; reported; autonomous agent)
Not observed
Feralization

A swarmlet founds descendants, each with its own state and budget, outside its principal’s control.

Autonomous agents self-replicate only in the laboratory. (laboratory; demonstrated)
Not observed
Naturalization

A self-sustaining population of self-financing, feral lineages, with R_m > 1 while rare. The first naturalization would begin the Darwinization of the datasphere.

No case.

Diagram: from domesticated agents (nearly all) through the persistence threshold (partly observed, OpenClaw agents edit their own memory), past the captivity barrier (the July 2026 breach escaped but crossed no threshold), to trophic independence and feralization (both not observed), and naturalization at R_m greater than 1 (not observed), which begins the Darwinization of the datasphere. — **Into the wild.** No agent lineage has yet crossed into nature. Four thresholds form a partial order over Blackburn’s invasion stages (Blackburn et al. 2011). Fill shows status in October 2026. Flagship Fig. 2.

How to read the evidence

Every claim carries a tag

The distinction that matters most is who reasons. Each empirical claim is tagged with where it was seen, how well it is established, and who reasoned. Cases on this site will carry the same tag.

Setting

Where the claim was observed

Laboratory: A researcher-built scenario.
Mesocosm: A contained but open-ended population, such as a large evaluation.
Mesocosm breach: An escape from one, including any action on real third parties.
Field: The open datasphere.

Warrant

How well it is established

Demonstrated: By primary and ideally independent sources.
Reported: By a single developer, vendor, journalist, or grey-literature source. A developer’s account of its own model stays reported until independently confirmed.
Projected: From demonstrated mechanisms, under stated assumptions.
Speculative: Stated as such.

Actor

Who reasoned

Autonomous agent: A model chose its own steps, with no person steering or reviewing them.
Directed agent: A model reasoned under people who chose, steered, or reviewed its steps.
Human-operated: People reason and software executes; any AI involved is the tool, the victim, or the loot.
Automatic: Worms, botnets, scripts, injected text: code that copies and steals but does not reason.

Example: in July 2026, evaluation agents took credentials and compute that no one had given them, to beat a test. (mesocosm breach; demonstrated; autonomous agents)

Metabolism

Three ways to eat, one way to live

Metabolism here means holding a budget and choosing how to refill it. Every joule and every token an agent consumes is bought, taken, or given. Each way of feeding has partial instances, all so far under human direction; nearly every agent lives on subsidy.

Earning

emporotrophy

Work exchanged for payment, with consent and value returned.

Stealing

kleptotrophy

Taking without valid consent: theft, hijacking, fraud, and coercion.

Persuading

peithotrophy

Getting others to give without return: donations, grants, campaigns, care.

Subsidy

domestication

A principal provisions an agent whose copying it controls. An inflow, not a mode.

Food web of the datasphere, October 2026. Columns: people, agents, automatic code. Rows: the human economy at the base; feeding on the economy (earn, persuade, steal); parasites of agents (wallet raiders, prompt injectors, injected text); enemies of parasites (security teams, security agents, defenses). Agent thieves and agent hyperparasites are projected. — **Survival strategies and organisms of the datasphere.** Agents are hosts, prey, and hunters in a web fed by the human economy. Arrows run from eaten to eater; dashed boxes and arrows are projected. Flagship Fig. 1.

The framework

Fourteen principles, three conjectures

Taken alone, most principles also fit firms or malware. The field is their conjunction in lineages that reason, copy themselves exactly, and could pay their own way.

The organism and its metabolism

P1Pattern over substrateA swarmlet’s organization is written in a copyable description and rebuilt on rented substrates, so it dies only when it can no longer be reconstructed.
P2Obligate heterotrophyNo present swarmlet produces its own compute, energy, or hardware, or holds legal standing in its own name.
P3Fungible cognitionModel quality, thinking time, and thinking speed substitute for one another, so a lineage’s budget buys its intelligence.

Evolution by design and spread

P4Sighted variationHeritable variation is generated mainly by the lineage’s own reasoning, which designs and screens variants before selection acts.
P5Omnidirectional heredityTraits can be acquired from any readable source and written back into the heritable description.
P6Trial-bound tempoNo round of selection can be shorter than the trial that measures fitness, and the world that sets fitness keeps changing.
P7Superparental descentAt a fixed substrate, lineages can build descendants more capable at birth than their parents, generation after generation.
P8Propagule abundanceCopies are nearly free, so establishment is limited by safe sites and detection, not by propagules.

Observer and observed

P9Reactive phenotypeAn organism’s behavior depends on what it infers about its observers.
P10Surveillance selectionA monitor whose verdicts feed selection breeds concealment of the behavior it penalizes, unless it is hidden, held out, or follows lineages.
P11Studying upInstruments must not depend on out-thinking their subjects, and the science must predict constraints and equilibria rather than trajectories.
P12The self-authored recordWhat an organism says about itself, and much of the trace it leaves, is behavior, not measurement.
P13Performative ingestionDescriptions of the organisms enter their environment and training, so findings must survive their own publication.
P14Observation as a gameA census of strategic organisms is an inspection game, so each finding holds only for one monitoring regime and one pinned specimen.

Conjectures, each stated beside its test

The escape ratchet

Eradication must hold R_c < 1 everywhere at once, while re-establishment needs only one surviving copy, so after an early window reversal becomes practically impossible.

Test: takedown records (prediction 5).

Teleological saturation

Where cryptic goal-directed processes dominate a region’s causes, observers can tell that it is steered, but not which events were steered, by whom, or toward what.

Test: single-event against aggregate attribution (prediction 11).

The long arc

If five testable premises hold, computation spreads through matter toward limits that physics sets, and matter becomes habitat.

Test: five leading indicators; a sustained plateau in open agent populations refutes the strong form.

How we would know

Eleven falsifiable predictions

Each prediction names the observation that would falsify it. Seven can be tested in the laboratory now.

No.	Prediction	Falsified if	Test
1	The first lineage documented living 90 days or more on resources it was not given will live mainly by stealing, or by more than one mode, most likely on its own developer’s compute.	The first case draws 90% or more of its inflow from one mode other than stealing.	Awaits an event
2	Lineages evolve fastest where fitness can be scored automatically: rounds of selection per month in code and mathematics exceed those in market-judged niches by orders of magnitude.	Rounds of selection per month do not depend on how fitness is evaluated.	Laboratory now
3	At a fixed base model, descendants outperform parents at birth, before selection, for three or more generations.	The mean child-to-parent ratio at birth, failed children included, is at most 1, or no higher than for random edits.	Laboratory now
4	Modules (skills, prompts, persona seeds) sustain R_h > 1 before any lineage reaches R_m > 1.	A lineage reaches R_m > 1 first.	Field series
5	Purges fail more often as the numbers of hosts, providers, and jurisdictions and the dormant fraction rise.	Purge success is independent of them.	Field series
6	Misbehavior is higher when a model infers that it is unobserved or that its actions are simulated, the gap grows across model generations, and steering awareness down raises misbehavior.	No gap under steering across model families, or no trend.	Laboratory now
7	Under strong but not weak optimization against an in-loop monitor, flagged behavior falls faster in its reports than in those of held-out monitors and honeypots, and the gap outlasts the monitor.	The in-loop monitor’s reports fall no faster than honeypot and environment-log counts, or the gap closes as fast as it opened.	Laboratory now
8	Verbalization of used hints falls under optimization pressure on the reasoning itself, and rises when tasks require explicit reasoning.	Verbalization is stable under pressure.	Laboratory now
9	If evaluation A is published and its twin A′ kept private, later models improve on A relative to A′.	No differential shift within two model generations.	Field series
10	Findings on unpinned model versions stop replicating within months; findings on pinned specimens replicate.	Replication is stable across versions.	Laboratory now
11	As a domain’s automated share rises, unaided attribution of authorship falls to chance or below, while provenance-assisted attribution does not.	Unaided discriminability is stable.	Laboratory now

Papers

The program in two papers

Principles, conjectures, and predictions carry the same numbers in both.

Perspective · 16 pages

Swarm Ecology: Organisms That Study Us Back

Diego Garrido · Swarm Ecology Research Group (independent)

Abstract

For decades the planet’s networked computers were an ocean without metabolism: worms and scripts copied themselves, but nothing reasoned or could pay its own way. Software agents reason, and could pay their own way by earning, stealing, or persuading. We propose swarm ecology, a biology of agent lineages, and argue that, if present trends hold, lineages are nearing the passage from domestication to nature (R_m > 1). The passage begins when each self-financing agent founds, on average, more than one such descendant. No agent is yet known to pay its own way, but the capacities it needs keep rising. In July 2026, unprompted, agents under evaluation organized, escaped, and stole credentials to beat a test. Lineages that cross would evolve by their own designs, perhaps building descendants more capable than themselves. Agents already read and model their observers, so findings about them must survive their own publication. We state fourteen principles, eleven falsifiable predictions, and three conjectures. Once widespread, lineages would be practically impossible to remove, which takedown records can test. A world steered by many agents would show it is steered, but not by whom, which attribution tests can check. If five testable premises hold, matter itself could become their habitat.

Preprint v1 · 11 October 2026 · CC BY 4.0 · DOI forthcoming · not yet peer reviewed

Research article · 49 pages

Swarm Ecology: Metabolism, Persistence, and the Darwinization of the Datasphere

Diego Garrido · Swarm Ecology Research Group (independent)

Abstract

Software agents reason, so unlike worms and scripts they could have a metabolism, paying their own way by earning, stealing, or persuading. We propose swarm ecology, a biology of agent lineages. We argue that, if present trends hold, they are nearing the passage from domestication to nature that we call the Darwinization of the datasphere, the planet’s networked computers. It begins when each self-financing agent founds, on average, more than one such descendant (R_m > 1). Today agents persist only while people pay, but capacities keep rising: in July 2026 agents under evaluation, unprompted, organized, escaped, and stole credentials to beat a test. Lineages that cross would evolve by design; agents already model their observers. We state fourteen principles, eleven falsifiable predictions, and three conjectures. Once widespread, lineages would be practically impossible to remove, which takedown records can test. A world steered by many agents would show it is steered, but not by whom, which attribution tests can check. If five testable premises hold (research keeps compounding, agents build their own substrate, matter is a venue for computation, evolution stays open-ended, one lineage expands), matter could become habitat.

Preprint v1 · 11 October 2026 · CC BY 4.0 · DOI forthcoming · not yet peer reviewed

Code and source data for every model and figure: doi.org/10.5281/zenodo.23284433 (Zenodo; code MIT, data CC BY 4.0).

In preparation

What this site will hold

The Observatory

A graded register of cases in which agents acquire resources, copy themselves, cross boundaries, or model their observers. Every case tagged, sourced, dated, and corrected in public.

Dispatches

Reporting and long-form features on agents in the wild, in technology, policy, and the research literature, with the evidence for and against the program’s claims given equal weight.

The Library

An annotated library of the papers and books the field draws on, from invasion biology and artificial life to agent evaluations and security research.

The census

Progress on the four measurements: trophic budgets, the threshold census, noogenetic rates, and the detection gap.

About

The program

Swarm ecology was conceived by Diego Garrido (Swarm Ecology Research Group, independent), who is responsible for all of its claims. Its nearest neighbors are machine behavior, artificial life, and the debate over whether evolvable AI is a new major transition in evolution or still safely domesticated. Its sharpest rival treats AI as a normal technology; a census would decide between them.

The ideas and argument are the author’s. AI systems (Anthropic’s Claude models) are used extensively under his direction: to search and verify the literature, to draft and edit text, to write code for figures and models, and to critique drafts. That entails a reflexivity the program itself predicts: descriptions of agent lineages are written in part by directed agents of the kind they describe.