Adaptive Modular Development builds towns by learning from each phase
Adaptive Modular Development (AMD) is a method for developing places in small, revisable phases. It is intended to connect urban planning theory, simulation, construction management, procurement, and robotics in software that can be developed by an individual or a small team and applied in regions that meet specific conditions. Its immediate objective is to put a working prototype into the world before pursuing completeness, so that a discussion can begin even where established players do not yet exist.
AMD works best where development can begin with limited interference
The method is intended for additional development on low-density land with little interference from existing buildings, and for declining regions where the industrial structure is empty and the cost of coordinating interests is low. These conditions are not incidental. They reduce the need for land assembly, extensive rights negotiations, and large-scale consensus procedures that make conventional urban development difficult to change once it has begun.
The case for a more incremental method comes from the difficulty of large urban projects. According to the cited research on megaprojects, cost overruns are commonplace, and the assumptions that once justified large-scale development are being weakened by population decline, modularization, and improvements in the accuracy of AI-based prediction. AMD therefore treats a town as a sequence of decisions rather than as a single commitment to a finished master plan.
Each phase should be independently viable and open to revision
AMD divides development into units that can be independently accounted for and independently checked. A delay or cost overrun in one phase should not automatically spread through the entire project. At the end of each phase, the situation is reassessed, and the project retains the right to change the next phase. Contracts and budgets should not make continuation automatic.
This principle changes what counts as a plan. Instead of fixing a long-term demand forecast, land-use arrangement, procurement package, and infrastructure scale at the outset, the project uses actual results to revise them. The simulator, procurement process, and construction management system are connected through the same data foundation, allowing performance to return to the plan quickly.
The hypothesis is that smaller phases can limit the damage from unusually large overruns. A phase that begins to expand can be stopped or corrected before it damages the whole project, and repeated phases may improve the accuracy of estimates. This effect is not verified. It depends on the phases being genuinely independent, on connection costs not becoming excessive, on optimistic bias being addressed, and on having clear criteria for stopping. Each phase would therefore need a maximum budget and pre-agreed stopping conditions.
AMD replaces irreversible commitments with a recurring decision loop
Compared with methods used to create large cities, AMD does not require a master plan fixed by a long-range forecast, advance infrastructure sized for a future peak population, a single large contract, fixed-use zoning, or a handoff between separate planning, design, construction, and procurement organizations. It also avoids permanent structures that are difficult to remove or move when the plan changes.
In their place, AMD requires criteria for deciding whether to stop, correct, or continue at the end of each phase; an independent financial judgment for each phase; immediate feedback from results into the plan; and a shared data foundation across disciplines. Because the initial development has no established local players, the method also needs a way to create initial demand and initial employment. A working prototype is part of that mechanism: it gives people something concrete to discuss, and it automates enough specialist knowledge for a small team to operate across disciplines.
Curitiba offers techniques to examine, not a success formula to copy
Curitiba is a relevant but limited comparison. It should not simply be described as a success story. Its early design was strong, while later updates exposed problems; the comparison is useful only when those two aspects are kept separate.
The elements close to AMD include a continuing planning organization established under the mayor in 1965, a basic plan approved in 1966, and gradual implementation over several decades. Curitiba integrated zoning with transport corridors and introduced bus rapid transit through a reorganized road structure rather than a subway. The system developed in stages, including the creation of the RIT in 1974 and tube stations and direct services in 1991. The approach associated with former mayor Lerner emphasized understandable actions and starting quickly, using low-cost small projects rather than a single massive investment. URBS connected the city and operators, while the bus system could theoretically adjust vehicle numbers and configurations to demand.
The limitations matter just as much. The claim that the system operates profitably without tax support does not describe the current situation: the cited secondary source reports 2024 operating costs of R$8.73 against user fares of approximately R$6.00, with the difference covered by a city subsidy. Fare recovery was profitable in the late 1990s and later declined. Public transport use also fell while car ownership and solo driving increased. About 60 percent of the articulated three-section buses were reported to have exceeded their service life, even though average daily transport reached 1.1 million in 2022. The system's ability to respond to demand has not fully addressed vehicle ageing and the post-pandemic downturn.
Curitiba's apparent success also depended on a long-running mayoral organization and a particular individual, so it is not a transferable formula. The city should not be used as a human-scale model for AMD's desired conditions of walkable areas, limited car access, proximity between work and residence, and everyday community interaction. Its road design assumes car movement; the industrial district reflects functional separation rather than work-residence proximity; and the cited material records serious living-standard problems. Pedestrian streets and civic facilities may be relevant, but their effect on everyday lingering and conversation has not been confirmed. Curitiba is therefore a source of techniques—continuity of the planning organization, gradual accumulation, and integrated land use and transport—not a guarantee that low initial costs or operating surpluses will last for decades.
A shared data layer makes the method operable for a small team
The software core is a shared data layer with collection agents. Rather than requiring contractors to adopt dedicated software, AI agents gather design, procurement, construction, and performance data from existing exchanges such as Excel files, LINE, email, invoice PDFs, and site photographs, then organize it under a common schema. Open formats preserve flexibility and reduce the burden of adoption.
The original documents, photographs, and messages should be retained as sources. Important amounts and quantities need a human verification route. Paper and oral information require an entry point through photographs or sensors, and contracts and prices require access controls. On top of the data layer, the project can provide an urban-layout simulator, a phase-level revenue model, a phase reassessment tool, construction and cost management, procurement management, and proposal-package output. These are different views of the same foundation rather than separate systems that contractors must all install.
Hardware should be modular because the plan will change
The hardware side includes factory-produced modular components that can be removed, moved, and expanded; construction robots and automated equipment; site sensors and drone surveying; distributed small-scale infrastructure added phase by phase; and mobile vendors or small public modules that create initial demand. Design and manufacturing should overlap rather than wait for a fully finalized design. Small lots, modular units, and removable components can limit the cost of rework while the next production lot incorporates field results.
This does not eliminate physical, legal, or human constraints. AI agents may help with document transfer, quantity estimates, design-change calculations, decision tracking, contract deadlines, and early detection of delays from photographs, sensors, and daily reports. They cannot remove statutory permitting periods, resident opposition, lawsuits, land negotiations, financing decisions, or the physical time required to manufacture and transport materials. They also introduce unresolved questions about mistaken judgments, contractual responsibility, contractor trust, and access to confidential prices and terms.
AMD is consequently a disciplined learning loop: start with a small independently judged phase, collect actual results through a common data foundation, revise both the plan and the hardware, and decide whether the next phase should stop, change, or proceed. Its promise remains a hypothesis, but its operating rules are designed to keep an unverified promise from becoming an irreversible commitment.