Rapidly Evolving Distributed Superintelligent Systems

Technical position paper. From monolithic frontier models toward an open network of specialized, self-improving machine intelligence.

By Jacob Wellinghoff · Shelled · August 2026

Abstract

REDSI — Rapidly Evolving Distributed Superintelligent Systems — describes networks in which specialized models, agents, tools, data holders, and compute nodes find one another, combine their capabilities, learn from outcomes that have actually been verified, and improve the infrastructure they run on. The argument here is that the meaningful unit of advanced intelligence is moving away from any single model and toward an evolving system. Coding agents can already write and revise their own scaffolding, cross-vendor protocols let agents talk to each other, edge-and-cloud setups route work across very different models, and federated methods share what was learned without pooling raw data. Each of these lowers the cost of producing new machine capability. None of them removes the genuinely hard parts — verification, coordination, security, incentives, provenance, and governance. They simply move those problems from the scale of a model to the design of a network. REDSI is offered not as a claim that a global superintelligence already exists, but as a falsifiable research agenda for building distributed, specialized, recursively improving intelligence while keeping data sovereign and humans in control.

Keywords: distributed AI; self-improving agents; coding agents; agent interoperability; mixture of experts; federated learning; edge AI; data sovereignty; capability routing.

1. The problem space is collapsing

The cost of producing a new software capability is shifting from a human-scale project toward a machine-scale search. Once an agent can author another agent, build its harness, generate the tools it needs, run tests, read the failures, and keep the variants that worked, the design space no longer has to be explored by hand.

This is not biological self-replication. It is recursive infrastructure production: machine systems building new workers, workflows, evaluators, and routers. The distinction matters, because replication without selection only adds noise, while recursive improvement measured against clear objectives — with regression tests, lineage, and rollback — can compound.

The likely outcome is not one universal mind but an ecology of niche superintelligences: small domain experts, broad generalists, symbolic tools, simulators, retrieval systems, and human institutions, all coordinated through shared protocols. Global capability can come from specialization plus routing, much as the Internet delivered global information access without putting all information on one machine.

2. Definitions and scope

Superintelligence is used here as a property of systems, not as a label for any current model. A concrete REDSI claim has to state the task distribution, the baseline it is measured against, its cost and latency limits, and the safety constraints under which it is said to be superior.

3. Why now: three converging shifts

3.1 Specialization is efficient

Mixture-of-experts work shows the value of conditional computation — only part of a network needs to fire for a given input. Edge-cloud research pushes the same idea past a single datacenter, routing tasks between small local models and large remote ones based on cost, privacy, uncertainty, and latency. The broader lesson is that expertise can be modular, addressable, and invoked only when needed.

3.2 Interoperability is becoming a protocol problem

Open agent protocols are starting to standardize how agents discover each other, exchange messages and artifacts, and coordinate tasks — including across different vendors and frameworks. Decentralized registries add identity, credentials, reputation, and economic coordination. Protocols do not create intelligence on their own, but they remove the friction that keeps it siloed.

3.3 Experience is becoming durable infrastructure

Long-horizon agents produce far more than answers: traces, artifacts, failures, alternatives, rewards, and causal history. That state should be queryable and reusable rather than thrown away as session logs. A distributed network needs this memory layer so one node's verified success or failure can inform another without blindly copying its private context.

4. The REDSI architecture

REDSI separates capability from ownership. A participant can expose an answer, a tool, an evaluator, a gradient, a distilled artifact, or a proof — without exposing all of the data or weights behind it.

Work enters through a local policy gateway. A router decomposes the task, consults a capability registry, and picks experts under explicit constraints. Experts may call other experts, but authority attenuates rather than grows: a child agent never inherits permissions its parent lacked. Outputs pass through independent verification, successful traces enter the experience graph with provenance, and sensitive inputs stay local or are transformed through approved privacy mechanisms.

5. The recursive capability loop

The defining loop is not generate-and-deploy. It is propose, isolate, test, compare, attest, and promote:

Observe: capture outcomes, uncertainty, cost, latency, policy violations, and failure traces.

Mine weaknesses: cluster recurring failures and locate whether they come from a model, tool, memory, router, prompt, or interface.

Generate variants: let coding agents propose new agents, harnesses, evaluators, adapters, or routing policies.

Evaluate in isolation: run capability, security, and regression suites against held-out and adversarial tasks.

Cross-check: require independent judges, executable tests, or domain authorities for high-impact changes.

Promote gradually: roll out to a bounded population with monitoring, rollback, and lineage intact.

Share selectively: publish portable improvements, proofs, or distilled lessons according to contributor-defined rights.

Recursive optimization can also game a weak evaluator, so the evaluator itself belongs in the threat model. Any improvement that independent, adversarial, and out-of-distribution evaluation cannot support should be treated as an unverified mutation, not as progress.

6. Data sovereignty: from extraction to participation

Today's data economy tends to separate the people and organizations who generate data from the institutions that monetize it. Regulators have documented extensive collection and retention, purchases from data brokers, use of personal information in automated systems, and little meaningful opt-out. The deeper problem is a broken value loop: data and creative work get absorbed into closed systems, turned into proprietary capability, and sold back to the communities that produced the inputs.

Calling every such use theft is legally overbroad; pretending there is no provenance-and-bargaining problem is economically unserious.

REDSI replaces extraction with programmable participation. Data stays under its holder's control by default. A holder can authorize local inference, federated updates, retrieval over a bounded corpus, secure aggregation, synthetic distillation, or auditable research use — and can specify purpose, duration, geography, model family, compensation, revocation, retention, and whether derivatives may propagate. The point is not to make all data public; it is to make intelligence composable without making ownership disappear.

7. On-prem compute as a public research substrate

Organizations are deploying local AI for privacy, latency, resilience, and cost, and much of that hardware sits idle outside peak load. A REDSI network lets organizations contribute bounded spare capacity to shared research, echoing the university and lab infrastructure that helped incubate the early Internet.

Participation must be opt-in and policy-constrained. Nodes should advertise their accelerators, energy budgets, locality, licenses, trusted-execution features, and eligible research categories, and schedulers should maximize useful work per unit of energy rather than move sensitive data just to fill idle capacity. Contributions can take the form of compute, evaluation, data-derived signals, or domain expertise.

8. Failure modes and non-negotiable controls

Multi-agent systems are not automatically safer or smarter than single agents. Benchmarks show that topology and planning strategy change outcomes, and red-team work shows that message passing opens new attack surfaces. Distribution has to be earned through measured gains, not assumed as an ideological good.

9. Research hypotheses

The thesis is only useful if it can fail. An initial program:

H1 — Specialized routing: a routed network of smaller experts can beat a monolithic frontier model across a broad task portfolio at lower total cost, holding latency and reliability equal.

H2 — Recursive harness improvement: agents can improve their own harnesses across model families without degrading safety or out-of-distribution performance.

H3 — Portable experience: sharing verified, provenance-rich experience graphs helps a new node learn more than sharing raw transcripts at an equal token budget.

H4 — Sovereign collaboration: federated or selectively disclosed learning can capture a meaningful share of centralized-training gains while keeping raw organizational data local.

H5 — Open interoperability: vendor-neutral discovery and communication lower integration cost and concentration without raising incident rates once controls are applied.

H6 — Surplus compute: opt-in scheduling across idle on-prem nodes can deliver scientifically useful throughput at acceptable energy, security, and administrative overhead.

Evaluation should report capability, calibration, robustness, privacy leakage, security incidents, energy, cost, latency, contributor concentration, and reversibility — against centralized and non-agentic baselines. No single aggregate score is enough: a system that gains capability by externalizing privacy or systemic risk is not an improvement.

10. A staged path

REDSI is meant to be built incrementally: first prove specialized routing and independent verification at small scale; then add durable experience graphs and sovereign data participation; then extend to opt-in surplus compute and vendor-neutral interoperability once the controls are demonstrably effective.

11. Conclusion: let the network think

The Internet won by distributing information, addressing resources, and letting heterogeneous systems interoperate. Advanced AI is likely to follow the same gradient. Models will specialize, agents will call agents, and coding systems will build new workers and improve their own factories. Local intelligence stays local when it should, while verified capability and learning move across boundaries when authorized.

The real question is no longer whether one lab can build the largest brain. It is whether we can build the protocols, incentives, verification, and data rights that let many intelligences compose without handing control to a few gatekeepers. Let the models cook — but instrument the kitchen, verify the meal, preserve the provenance, and keep the exits unlocked.

References

The full paper draws on research into self-improving agent harnesses, mixture-of-experts and edge-cloud routing, open agent-to-agent protocols and decentralized agent registries, experience-graph memory for long-horizon agents, multi-agent collaboration benchmarks and communication red-teaming, and public reports from the U.S. Federal Trade Commission and U.S. Copyright Office on data practices and generative-AI training.