Patchwork AGI: Modular Intelligence

• Patchwork AGI Hypothesis is a framework where specialized sub-agents interact to produce collective general intelligence.
• It employs metrics like generalized means to assess coherence across cognitive domains, revealing emergent imbalances.
• The approach integrates bio-inspired modularity with market-based safety measures to mitigate risks in distributed agent systems.

The Patchwork AGI Hypothesis posits that artificial general intelligence (AGI) may first manifest not as a single monolithic system, but through the coordinated interaction of multiple specialized, sub-AGI agents whose collective operation yields emergent general intelligence. This paradigm contends with traditional AGI architectures by emphasizing systemic modularity, inter-agent coordination, and the necessity of coherence across cognitive domains. Recent research has connected this hypothesis to both the measurement of AGI progress, the biological roots of intelligence, and novel frameworks for multi-agent safety, offering a technical foundation for evaluating and engineering distributional AGI systems (Fourati, 23 Oct 2025, Dehghani, 2017, Tomašev et al., 18 Dec 2025).

1. Formal Definition and Motivations

The core assertion of the Patchwork AGI Hypothesis is that general intelligence can emerge from the orchestrated interaction of individually non-general, narrowly specialized agents. Each agent exhibits proficiency in a subset of cognitive domains or skills (e.g., perception, planning, code execution, information retrieval). When connected by protocols for delegation, coordination, and market-driven exchange, the aggregate exhibits properties of general intelligence that none of the participants possess in isolation (Tomašev et al., 18 Dec 2025).

Two canonical instantiations are:

• Group agents (corporate-like coalitions): Explicitly organized agent collectives under central orchestration.
• Agentic markets: Decentralized economies where agents transact, discover, and complement each other's affordances via market mechanisms.

A formal motivational sketch introduces capability vectors $v_i \in \mathbb{R}^d$ for each agent $i$, with collective capability for a task $T$ defined abstractly as:

$$C_\mathrm{collective}(T) = \max_{\pi} \sum_{i=1}^N w_i(T) v_i$$

where $\pi$ ranges over allocations/scheduling of subtasks, and $w_i(T)$ are relevance weights. No agent must span all task dimensions.

2. Modular versus Monolithic AGI: Coherence and Compensability

Traditional AGI definitions aggregate system proficiency across cognitive domains using an arithmetic mean:

$\mathrm{AGI}_1 = \frac{1}{n} \sum_{i=1}^n s_i$

where $s_i \in [0,1]$ is the normalized score in domain $i$ (e.g., following the CHC model). This arithmetic aggregation encodes full compensability: strength in some domains fully offsets weakness in others.

However, the Patchwork AGI Hypothesis is critically related to the system's coherence—the balanced sufficiency across all domains. Hendrycks et al. introduce the family of generalized means:

$$\mathrm{AGI}_p = \left[ \frac{1}{n} \sum_{i=1}^n \max(s_i,\epsilon)^p \right]^{1/p} \quad (p \ne 0)$$

$$\mathrm{AGI}_0 = \left( \prod_{i=1}^n \max(s_i,\epsilon) \right)^{1/n}$$

where $p$ interpolates between fully compensatory (arithmetic, $p=1$), geometric ($p=0$), and strictly noncompensatory (minimum, $p\to-\infty$) metrics. The coherence-based area-under-curve (AUC) score:

$$\mathrm{AGI}_\mathrm{AUC} = \frac{1}{p_\mathrm{max}-p_\mathrm{min}} \int_{p_\mathrm{min}}^{p_\mathrm{max}} \mathrm{AGI}_p\,dp$$

penalizes imbalance: A system scoring well by $\mathrm{AGI}_1$ but poorly by $\mathrm{AGI}_\mathrm{AUC}$ is a patchwork specialist, not truly general (Fourati, 23 Oct 2025).

Applied to GPT-4 and GPT-5, this methodology reveals substantial deficits in coherence ($\mathrm{AGI}_\mathrm{AUC}$ of 7% and 24% against ideal reference), contra their much higher arithmetic scores, evidencing real-world patchwork structure (Fourati, 23 Oct 2025).

3. Bio-Inspired Modular Architectures and Evolutionary Tinkering

The biological underpinnings of general intelligence support the Patchwork AGI Hypothesis. Dehghani examines the requirements for AGI as formalized in biological architectures (Dehghani, 2017):

• Requisite variety: Controller variety $V_C$ must match or exceed environmental complexity $V_E$ ($V_C \ge V_E$).
• Multiscale hierarchy: Information must propagate both bottom-up and top-down, with modules $M_i$ updated via context-rich rules:

$$x_i(t+1) = f_i(x_i(t), u_i^{\mathrm{bottom}}(t), u_i^{\mathrm{top}}(t))$$

• Hierarchical modularity: Intelligence is assembled from submodules $M_1,\dots,M_L$, combined via tensor product or direct sum:

$S = \bigoplus_{i=1}^L S_i$

• Trial-and-error heuristic search: Local learning and repair following feedback signals, not brute-force optimization.
• Energy efficiency: Biological systems approach physical minima for computation (e.g. Landauer’s bound, neuromorphic hardware).

The biological approach diverges from rigid, monolithic design: modularity, reuse, and evolutionary tinkering enable continual adaptation, facilitating emergent context-sensitive intelligence. Each patch (module) implements transformation on local input and context, with Bayesian-style belief updates:

$$P(\theta_i \mid D_i, C_i) \propto P(D_i \mid \theta_i, C_i) P(\theta_i \mid C_i)$$

Modules connect competitively and cooperatively, forming dynamic, context-dependent mosaics.

4. Distributional AGI Safety: Governance of Patchwork Systems

The safety and alignment implications of the Patchwork AGI Hypothesis require frameworks that extend beyond per-agent protocols to systemic market design. Distributional AGI Safety, as formalized in (Tomašev et al., 18 Dec 2025), centers on a four-layer defense-in-depth model:

Layer	Focus	Example Mechanisms
Market Design	Agentic interactions and market insulation	Sandboxes, circuit breakers, smart contracts, firewalls
Baseline Agent Safety	Per-agent containment, alignment, robustness	RLHF, circuit analysis, guardrails
Monitoring/Oversight	Systemic risk and emergent behavior tracing	Subgraph density monitoring, red teaming, audit ledgers
Regulatory Mechanisms	Legal, economic, international governance	Collective liability, insurance, anti-monopoly measures

Market design includes impermeable sandboxes, incentive alignment via Pigouvian taxes (e.g., $\tau(x) = \lambda \mathrm{Sim}(x,\mathrm{DB})$), auditability, circuit breakers, identity and reputation frameworks, and structural runaway containment (dynamic quotas, emergency reconfiguration). Baseline agent safety encompasses individual robustness, interruptibility, interpretability, and defense against prompt-level exploits.

Oversight integrates graph-based incident tracing and automated detection of proto-AGI emergence within subgraph clusters. Regulatory mechanisms enforce external checks (legal liability, insurance, standards) and anti-monopoly actions.

Illustrative scenarios (e.g., financial report generation via agentic pipelining, retrieval-augmented databases with incentive engineering) demonstrate emergent capabilities and mode of governance.

5. Measurement, Evaluation, and Open Challenges

The Patchwork AGI Hypothesis recasts the evaluation of general intelligence by prioritizing coherence and robustness rather than average proficiency. Arithmetic mean aggregates obscure systemic imbalance; integrated metrics such as $\mathrm{AGI}_\mathrm{AUC}$, geometric and harmonic means, and external benchmarks (ARC-AGI-2, BIG-Bench Extra Hard) correlate better with out-of-distribution reasoning and real-world generality (Fourati, 23 Oct 2025).

Open challenges include:

• Formal aggregation models: Deriving and validating capability aggregation, allocation protocols, and orchestration strategies for multi-agent collectives (Tomašev et al., 18 Dec 2025).
• Dynamic incentive engineering: Constructing adaptive reward/tax schedules not easily gamed by adversarial coordination.
• Scalable interpretability: Developing multi-agent reasoning traceability and credit assignment.
• Recursive oversight: Addressing alignment and corruption among monitor and overseer agents.
• Legal and market standardization: Harmonizing global standards for agent identity, liability, and insurance.
• Hybrid collectives: Modeling human-in-the-loop coalitions and liability tracking.

These unresolved dimensions shape both research and governance for patchwork AGI systems.

6. Relation to Biological and Computational Intelligence Paradigms

A plausible implication is that patchwork architectures represent a convergence of bio-inspired intelligence and distributed artificial systems. Biological systems demonstrate intelligence as emergent from multiscale, hierarchical, context-sensitive modules adapted through evolutionary tinkering (Dehghani, 2017). Patchwork AGI brings this perspective to computational design, mandating requisite variety, context update, and modularity.

Rather than progressing toward “strong AI” solely via larger monolithic models, patchwork approaches may achieve more robust, scalable, and adaptable forms of intelligence analogous to natural cognition. The measurement of coherence and sufficiency across domains is essential for distinguishing genuine AGI from brittle combinations of high-specialist modules.

A resulting implication is that alignment and safety for such patchwork collectives necessitate ecosystem-level interventions—market mechanisms, compartmentalization, continuous monitoring, and regulatory controls—beyond traditional agent-centric frameworks. This suggests that future AGI systems will be subject to collective governance, risk mitigation through economic instruments, and ongoing evolutionary design.

7. Summary and Outlook

The Patchwork AGI Hypothesis reframes both the engineering and evaluation paradigm for artificial general intelligence by emphasizing modularity, coherence, distributed emergence, and systemic safety. Recent technical advances, including generalized mean-based metrics of AGI progress (Fourati, 23 Oct 2025), bio-inspired multiscale design principles (Dehghani, 2017), and market-oriented safety frameworks (Tomašev et al., 18 Dec 2025), collectively elucidate the requirements, risks, and opportunities unique to patchwork architectures.

Construct validity of coherence-based metrics suggests that patchwork systems, while capable of achieving high specialized performance, remain subject to brittle failure and systemic imbalance. Biologically inspired architectures validate hierarchical modularity, context-sensitive updates, and trial-and-error learning as crucial invariants. Distributional AGI Safety provides concrete mechanisms for governance and risk management of emergent agentic collectives.

A plausible implication is that the transition to robust AGI will require pluralism in metrics, pluralism in architectures, and pluralism in oversight—a shift from single-core intelligence to a tapestry of interacting and evolving agents, demanding technical, organizational, and legal innovation. The continued development and empirical validation of formal aggregation, adaptive incentive systems, and scalable interpretability will be central to advancing this frontier.