Urban systems and distributed AI systems share a deep structural homology: both are composed of many agents with partial information making local decisions that aggregate into emergent global behavior. The failures of urban planning and the failures of distributed AI systems trace to the same root causes — information asymmetry, misaligned incentives, missing feedback loops, and governance structures that can't adapt faster than their environment changes.
The dissertation argues that the most important insights in 20th century urban theory map directly onto the design of resilient, adaptive, distributed agent systems — and vice versa. Cities are the oldest continuous experiment in distributed cognition at scale. We should learn from them.
---
Friedrich Hayek's critique of central planning applies with full force to urban planning: no central authority can aggregate the dispersed, local, tacit knowledge that millions of urban actors possess. Jane Jacobs demonstrated this empirically — the "eyes on the street" that make neighborhoods safe, the informal economy that sustains communities, the knowledge of which store is actually a community hub — none of this is legible to planners working from maps and statistics.
This is not an argument against planning. It's an argument about the epistemology of planning. The planner's advantage is scale and coordination. The resident's advantage is granular, contextual knowledge. Effective urban governance systems are those that find mechanisms to combine both: participatory planning that genuinely incorporates local knowledge (not theater), adaptive zoning that can respond to feedback, iterative implementation that learns from deployment.
The AI parallel: Centralized models have the same problem. A single model trained on aggregate data cannot possess the contextual, situational knowledge that emerges from being embedded in a specific environment. Distributed systems that learn from local context while coordinating globally are more robust — not because distribution is inherently good, but because it preserves the knowledge that lives at the edge.
---
Jane Jacobs identified the core pathology of urban renewal: it severed feedback loops. By destroying existing neighborhoods and replacing them with planned developments, planners eliminated the feedback mechanisms — occupancy rates, street activity, business survival — that would have told them whether their designs were working. They built without learning.
Successful urban systems maintain tight feedback loops between intervention and outcome:
The dysfunctional pattern: plan comprehensively, implement fully, evaluate never. Urban renewal, highway construction, urban renewal — all followed this pattern. Billions spent on interventions that made conditions worse, with no mechanism for detecting the failure until communities had been destroyed.
The AI parallel: Open-loop dispatch — systems that fire actions without monitoring outcomes — are the urban renewal of AI. TILE OBSERVE/EVALUATE/ADAPT is the architectural answer: every action logged, outcomes tracked, policy updated. The cortex doesn't plan once and execute forever; it accumulates outcome data and adjusts routing based on what actually worked.
---
Cities operate across radically different timescales simultaneously:
Governance failures often arise from timescale mismatch: electoral cycles (4 years) are too short for infrastructure investment (30-100 year payback), too long for operational response (days). Robert Moses was able to build so much partly because he operated at a timescale outside democratic accountability — he outlasted the politicians who appointed him.
Resilient urban governance separates decisions by timescale:
The AI parallel: Neuromod implements this directly — FOCUS/EXPLORE/ALERT/RESTORE modes operate at different timescales and adjust the system's behavior without requiring human intervention for every parameter change. The cortex tick (30s) handles reflexes; TILE EVALUATE (daily) handles medium-term policy; the skill crystallizer (200-session scan) handles long-term behavioral patterns.
---
Scale-free networks (few hubs, many spokes) are efficient but fragile — remove a hub and the network collapses. Random networks are robust but inefficient. Urban street grids that work best are neither: they're small-world networks with local clustering (neighborhoods) and long-range connections (arterials and highways). Jane Jacobs' "mixed primary uses" — the diversity of anchors that generate foot traffic at different times — is a prescription for small-world network topology in land use.
The failed urban renewal projects were scale-free in land use: one large anchor (housing project, shopping mall, stadium) surrounded by dependency. When the anchor failed or moved, the entire district collapsed. Successful urban districts are resilient because their diversity means no single failure is catastrophic.
The AI parallel: Thompson Sampling with multiple nodes (Pi/iMac/COZ) implements small-world resilience. No single task type depends on a single node. Browser tasks default to iMac but Pi can handle them. File operations default to Pi but iMac can handle them. The routing policy learns from outcomes, and the system degrades gracefully rather than failing catastrophically when a node is unavailable.
---
Cities generate massive value — agglomeration economies, network effects, rising land values from collective investment. That value is systematically captured by landowners who did nothing to create it, while the workers, businesses, and public investments that created the value capture little. This is Henry George's insight from 1879. It remains unimplemented because the political economy works against it: landowners have concentrated interests and high political power; renters and future residents have diffuse interests and low political power.
Community land trusts, participatory budgeting, and land value taxation are all mechanisms for redistributing collectively-created value back to the collective. They succeed in proportion to their ability to shift the political economy — making the beneficiaries of redistribution into a coherent political constituency.
The AI parallel: The knowledge graph and memory provenance system is the value capture layer for AI cognition. Every session generates insights, patterns, skills. Without structured capture, that value dissipates — the system has to re-learn the same lessons across sessions. The TILE crystallizer captures recurring tool-call patterns as proposed skills. The memory fact system captures structured knowledge with provenance. These are mechanisms for retaining the value that cognition produces rather than letting it leak away on context reset.
---
From urban planning theory, five design principles for distributed cognition systems:
1. Preserve feedback loops. Never dispatch without monitoring outcomes. Every action that matters should be observable. This is the lesson of urban renewal — building without learning produces catastrophe at scale.
2. Maintain timescale separation. Don't collapse reflexes and strategic decisions into the same loop. Fast systems should be fast. Slow systems should have long horizons and be insulated from fast-timescale noise.
3. Distribute knowledge, coordinate behavior. Each node should have access to local knowledge unavailable to other nodes. Coordination should be lightweight — shared state files, message queues, weekly policy updates — not centralized control.
4. Design for resilience over efficiency. Scale-free networks optimize for efficiency and fail catastrophically. Small-world networks accept some inefficiency in exchange for graceful degradation. The routing policy that uses Thompson Sampling with multiple capable nodes is deliberately not optimally efficient — it explores alternatives even when one is clearly better, maintaining fallback capacity.
5. Capture value from cognition. The intelligence of a system that doesn't learn from its own operation is being wasted. Every session, every wake, every routing decision is data. The question is whether the infrastructure exists to capture and act on it. Building that infrastructure is the prerequisite for genuine long-term improvement.
The city that works is not the city that was perfectly planned. It's the city that built institutions capable of learning from what it built.