Rapidly Evolving Distributed Superintelligent Systems

TECHNICAL POSITION PAPER

From monolithic frontier models to an open network of specialized, self-improving machine intelligence.

Jacob Wellinghoff · WebASI · August 2026

Abstract

This paper proposes Rapidly Evolving Distributed Superintelligent Systems (REDSI): networks of specialized models, agents, tools, data holders, and compute nodes that discover one another, compose capabilities, learn from verified outcomes, and improve their own operating infrastructure. The central thesis is that the relevant unit of advanced intelligence is shifting from a single model to an evolving system. Coding agents can already synthesize and revise harnesses; agent protocols can bridge vendors; edge-cloud systems can route work across heterogeneous models; and federated techniques can extract shared learning without centralizing raw data. Together these developments shrink the cost of creating new machine capabilities. They do not eliminate the hard problems of verification, coordination, security, incentives, provenance, and governance. They relocate those problems from model scale to network design. REDSI is therefore presented not as a claim that global superintelligence already exists, but as a falsifiable systems research agenda for building distributed, specialized, recursively improving intelligence while preserving data sovereignty and human control.

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

1. Thesis: the problem space is collapsing

The cost of producing a new software capability is falling from a human-scale project toward a machine-scale search problem. When an agent can write another agent, synthesize its harness, generate tools, run tests, inspect failures, and retain successful variants, the design space no longer has to be traversed manually. Recent work on AutoHarness and Self-Harness demonstrates the early form of this transition: models can generate or improve executable scaffolding around themselves, and validated harness changes can materially improve held-out task performance [1, 2].

This is not literal biological self-replication. It is recursive infrastructure production: machine systems creating new machine workers, workflows, evaluators, routers, and factories. The distinction matters. Replication without selection creates noise; recursive improvement with measurable objectives, regression testing, lineage, and rollback can create compounding capability.

The resulting future is unlikely to be one universal brain. It is more plausibly an ecology of niche superintelligences: compact domain experts, large generalists, symbolic tools, simulators, retrieval systems, and human institutions coordinated through shared protocols. A legal agent need not contain molecular biology. A protein-design system need not master procurement. Global capability can emerge from specialization plus routing, just as the Internet produced global information access without placing all information on one computer.

2. Definitions and scope

The word superintelligence is used here as a systems property, not a marketing label for any current model. A REDSI claim must identify the task distribution, baseline population, cost and latency limits, and safety constraints under which the network is said to be superior.

3. Why intelligence will become distributed

3.1 Specialization is economically and computationally attractive

Mixture-of-experts research shows the value of conditional computation: only a subset of specialized capacity must be activated for a given input [3]. Edge-cloud research extends the idea beyond a single datacenter by routing tasks among local small models and remote large models according to cost, privacy, uncertainty, and latency [4]. The architectural implication is larger than MoE: expertise can be modular, addressable, and invoked only when needed.

3.2 Interoperability is becoming a protocol problem

Open agent protocols are beginning to standardize discovery, messaging, artifacts, and task coordination. Google's Agent2Agent protocol explicitly targets collaboration among agents built by different vendors and frameworks, while work on decentralized agent registries adds identity, credentials, reputation, and economic coordination [5, 6]. Protocols alone do not create intelligence, but they reduce the friction that keeps intelligence siloed.

3.3 Experience is becoming durable infrastructure

Long-horizon agents generate more than answers. They produce execution traces, artifacts, failures, alternatives, rewards, and causal lineage. Experience-graph research argues that this state should be queryable and reusable rather than discarded as session logs [7]. A distributed intelligence network will require this memory layer so that one node's verified failure or success can inform another node without blindly copying its private context.

4. The REDSI architecture

REDSI separates capability from ownership. A participant may expose an answer, a tool, an evaluator, a gradient, a distilled artifact, or a proof without exposing all underlying data or weights.

A task enters through a local policy gateway. The router decomposes it, queries a capability registry, and selects experts under explicit constraints. Experts may call other experts, but authority attenuates rather than expands: a child agent cannot inherit permissions its parent did not possess. Outputs pass through independent verification. Successful traces enter the experience graph with provenance, while sensitive inputs remain local or are transformed through approved privacy mechanisms.

5. The recursive capability loop

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

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

Mine weaknesses: cluster recurring failure modes and identify whether they originate in a model, tool, memory, router, prompt, or interface.

Generate variants: allow coding agents to 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: deploy to a bounded population with monitoring, rollback, and lineage intact.

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

Self-Harness reports meaningful held-out gains from this style of loop, but recursive optimization can also exploit weak evaluators. The evaluator therefore becomes part of the threat model. A REDSI system should treat any improvement unsupported by independent, adversarial, and out-of-distribution evaluation as an unverified mutation, not progress [1, 8].

6. Data sovereignty: from extraction to participation

Today's data economy often separates the people and organizations that generate data from the institutions that monetize it. The U.S. Federal Trade Commission found that major social and video platforms collected and retained extensive data, obtained information from data brokers, used personal information in automated systems including AI, and frequently offered little or no meaningful opt-out [9]. The U.S. Copyright Office has likewise concluded that generative-AI training can implicate copyright owners' rights, while emphasizing that fair-use outcomes depend on the facts and remain for courts to decide [10].

The policy failure is not merely privacy loss. It is a broken value loop: data and creative work may be absorbed into closed systems, converted into proprietary capability, and sold back to the communities that produced the inputs. Calling every use theft is legally overbroad; pretending the provenance and bargaining problem does not exist is economically unserious.

REDSI replaces extraction with programmable participation. Data remains under the control of its holder by default. A holder may authorize local inference, federated updates, retrieval over a bounded corpus, secure aggregation, synthetic distillation, or auditable research use. Rights can specify purpose, duration, geography, model family, compensation, revocation, retention, and whether derivatives may propagate. The goal 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

Businesses are deploying local AI for privacy, latency, resilience, and cost control. Much of that compute will be idle outside peak workloads. A REDSI network can allow organizations to allocate bounded spare capacity to shared research, similar in spirit to the university and laboratory infrastructure that helped incubate the early Internet. Edge-cloud collaboration research already treats heterogeneous local and remote compute as a joint scheduling problem [4].

Participation must be opt-in and policy constrained. Nodes should advertise available accelerators, energy budgets, locality, model licenses, trusted execution features, and eligible research categories. Schedulers should optimize useful work per unit of energy and avoid moving sensitive data merely to fill idle capacity. Organizations may contribute compute, evaluation, data-derived signals, or domain expertise; not every participant must contribute the same resource.

8. Failure modes and non-negotiable controls

Multi-agent systems are not automatically safer or smarter than single agents. MultiAgentBench shows that topology and planning strategy affect outcomes, and red-team work demonstrates that message-based collaboration creates new attack surfaces [8, 11]. Distribution must therefore be earned through measured gains, not assumed as an ideological good.

9. Research hypotheses and evaluation program

The REDSI thesis becomes useful only when it can fail. The following hypotheses define an initial program:

H1 - Specialized routing: A routed network of smaller experts can exceed a monolithic frontier model on a broad task portfolio at lower total cost, subject to equal latency and reliability constraints.

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 improves new-node learning more than sharing unstructured transcripts at an equal token budget.

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

H5 - Open interoperability: Vendor-neutral agent discovery and communication reduce integration cost and concentration without increasing incident rates after controls are applied.

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

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

10. A staged path to global distributed intelligence

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

11. Conclusion: let the network think

The Internet won by distributing information, addressing resources, and allowing heterogeneous systems to interoperate. Advanced AI is likely to follow the same pressure gradient. Models will specialize. Agents will call agents. Coding systems will create new workers and improve their own factories. Local intelligence will remain local when it should, while verified capabilities and learning move across boundaries when authorized.

The strategic question is no longer whether one laboratory can build the largest brain. It is whether society can build the protocols, incentives, verification systems, and data rights that allow many intelligences to compose without surrendering control to a few gatekeepers. REDSI is a proposal for that stack: distributed intelligence, recursive improvement, open communication, sovereign data, and shared compute - with every increase in capability paired to provenance, tests, permissions, and rollback.

The age of niche superintelligence is not guaranteed. It is an engineering choice. Let the models cook - but instrument the kitchen, verify the meal, preserve the recipe's provenance, and keep the exits unlocked.

References

[1] X. Wang et al. Self-Harness: Harnesses That Improve Themselves. arXiv:2606.09498, 2026.

[2] X. Lou et al. AutoHarness: Improving LLM Agents by Automatically Synthesizing a Code Harness. arXiv:2603.03329, 2026.

[3] S. Mu et al. A Survey on Mixture of Experts. arXiv:2407.06204, 2024.

[4] Collaborative Inference and Learning between Edge SLMs and Cloud LLMs: A Survey of Algorithms, Execution, and Open Challenges. arXiv:2507.16731, 2025.

[5] Google Cloud. Announcing the Agent2Agent Protocol (A2A). 2025.

[6] S. B. Balija et al. The Trust Fabric: Decentralized Interoperability and Economic Coordination for the Agentic Web. arXiv:2507.07901, 2025.

[7] Experience Graphs: The Data Foundation for Self-Improving Agents. arXiv:2606.29823, 2026.

[8] K. Zhu et al. MultiAgentBench: Evaluating the Collaboration and Competition of LLM Agents. ACL 2025.

[9] U.S. Federal Trade Commission. Staff Report on Social Media and Video Streaming Data Practices. 2024.

[10] U.S. Copyright Office. Copyright and Artificial Intelligence, Part 3: Generative AI Training. 2026.

[11] P. He et al. Red-Teaming LLM Multi-Agent Systems via Communication Attacks. Findings of ACL 2025.

[12] Efficient Multi-Agent Collaboration with Tool Use for Online Reasoning. Findings of NAACL 2025.