5 Software Development Constraints That AI Has Eliminated

Published: February 17, 2026 | PRJ-02 Content Search Intent: Informational Keywords: AI changed software development, AI removes development constraints, how AI transforms coding

---

The Setup

Before 2023, software development operated under five constraints so fundamental that practitioners stopped seeing them as constraints. They were treated as physics --- immutable laws that every methodology had to accommodate. Scrum, Lean Startup, EOS, SAFe, Kanban --- every major framework was a rational response to these five limitations. The frameworks were not wrong. They were precisely calibrated to a constraint environment that no longer universally applies.

The conventional narrative around AI in software development focuses on productivity gains. GitHub surveys report that developers using Copilot complete tasks faster. McKinsey publishes figures on developer velocity improvements. Bain tracks engineering hours saved. These are real numbers, but they describe the wrong phenomenon. They measure AI as a tool --- a faster hammer hitting the same nails. The more fundamental shift is that AI has removed the constraints that defined what nails could be hit in the first place.

The difference between "tool" and "environment change" is not semantic. It determines whether organizations need minor workflow updates or fundamental methodology redesign. If AI is a faster tool, Scrum needs a velocity adjustment. If AI has removed the constraints Scrum was designed to manage, Scrum needs to be re-evaluated from first principles. The data supports the latter interpretation.

What the Data Shows

McKinsey's 2023 report, The Economic Potential of Generative AI, identifies software development as one of the sectors most impacted by generative AI. Their analysis estimates that generative AI could automate a significant portion of software development activities --- not just code writing, but debugging, documentation, testing, and code comprehension. The report positions these gains not as marginal improvements but as structural changes to how engineering organizations allocate human attention.

GitHub's survey data on AI in software development provides the practitioner-level view. Their research indicates that developers using AI assistants report not just speed improvements but fundamental changes in workflow patterns. Developers report spending less time on boilerplate, less time searching documentation, less time context-switching between reference material and code. The pattern is not "same workflow, faster." It is "different workflow entirely."

Bain & Company's data on AI engineering productivity reinforces the structural interpretation. Their analysis across engineering organizations shows that the gains from AI are not uniformly distributed across tasks. Routine, pattern-based work sees the largest acceleration. Novel, judgment-intensive work sees smaller but still meaningful gains. The implication: AI does not uniformly speed up development. It selectively eliminates certain categories of work, changing the ratio of human attention devoted to judgment versus execution.

Within the CEM (Compounding Execution Method) framework, developed and validated by Michael George Keating through the production of 596,903 lines of code in 10 systems (October 2025 -- February 2026), these external observations are operationalized as five specific constraint removals. The CEM foundation papers (F0 and F1) document each constraint, its historical impact on methodology design, and the mechanism by which AI dissolved it.

Here are the five constraints and how each was eliminated.

Constraint 1: Context switching was expensive --- AI made it cheap. Gloria Mark's research documented a 23-minute average resumption time after task interruption. Every methodology responded by protecting focus: Scrum with sprints, EOS with Rocks, Lean Startup with single-hypothesis cycles. AI dissolves this constraint through context preservation. Conversation history, project documentation, and code context persist in AI systems. When an operator returns to a project after working on another, the AI provides continuity: the state of the work, the outstanding issues, the next steps. The CEM portfolio demonstrates the result: up to 5 projects active on a single day, 132 commits across 4 projects on the peak parallel day (October 21, 2025), with no observable velocity degradation from switching. The 23-minute tax dropped to near zero.

Constraint 2: Expertise was scarce and localized --- AI made it available on demand. A backend developer could not do frontend work without significant ramp-up. A designer could not write database queries. Knowledge lived in individual heads and transferred slowly. Transaction Cost Economics (Williamson, 1975) formalized this: acquiring specialized knowledge through the market is expensive, so firms internalize it through hiring specialists, which creates coordination overhead. AI provides domain knowledge on demand. The CEM portfolio spans PHP, JavaScript, Golang, SQL, HTML/CSS, Laravel, Vue.js, Tailwind, and custom Go frameworks --- plus integrations with Stripe, Twilio, SendGrid, Shopify, and affiliate networks. No single operator possesses native expertise across all these domains. AI provided the expertise in context, without coordination overhead, without specialist bottlenecks.

Constraint 3: Learning required time away from execution --- AI merged them. Sweller's Cognitive Load Theory (1988) established that learning and performing compete for the same cognitive resources. You could not simultaneously master a new framework and ship production code. This forced sequential approaches: learn, then build. AI collapses the separation. The CEM operator learned Golang while building PRJ-04 --- a 5-day MVP with 100% primary commits in a language the operator had never used. The operator learned webhook architecture while building PRJ-01's ingestion system. The progression is visible in the data: October 2025 showed 31--43% primary operator commits (heavy AI support during learning). By January 2026, primary operator commits reached 96--100% (full independence at higher velocity). Traditional path to equivalent capability: 4-year CS degree plus 5--7 years professional experience. CEM path: 4 months.

Constraint 4: Building was expensive --- AI made it cheap. MVP development through traditional channels cost $50,000 to $250,000. Even lean approaches required weeks of developer time at $150--$250/hour. This expense justified the entire Lean Startup methodology: validate before building, because building wastes resources if the idea is wrong. When AI collapses building costs, the Lean Startup sequence inverts. Building becomes validation. The CEM portfolio demonstrates this inversion directly: PRJ-05 reached functional MVP in 4 days with zero product defects. PRJ-03 reached MVP in 4 days in a new vertical. PRJ-04 reached MVP in 5 days using a new programming language. Traditional validation for these projects would have required weeks of customer interviews, competitive analysis, and requirements documentation. Building functional systems was faster than researching whether to build them.

Constraint 5: Coordination overhead scaled with team size --- AI eliminated the team requirement. Brooks's Law (1975) observed that communication pathways scale as n(n-1)/2. A team of 5 has 10 channels. A team of 10 has 45. A team of 20 has 190. Every methodology addressed this: Scrum capped teams at 7--9, SAFe created teams-of-teams hierarchies, EOS structured accountability charts. When one operator with AI achieves output previously requiring teams, coordination overhead drops to zero. The CEM portfolio's flagship project, PRJ-01, comprises 194,954 lines of code, 135 database tables, and 20 external integrations (12 inbound, 8 outbound). Comparable platforms --- Segment, mParticle, Tealium --- required venture funding, multi-year development timelines, and full engineering teams. PRJ-01 was built by one operator in four months. The Carta Solo Founders Report (2025) documents the macro trend: solo-founded startups rose from 17% in 2017 to 36% in 2024.

How It Works

The five constraint removals do not operate independently. They compound. Cheap context switching enables parallel projects. On-demand expertise enables work across those parallel projects without specialist bottlenecks. Merged learning and execution means each project builds operator capability that transfers to the next. Cheap building means each project can be validated through deployment rather than research. Zero coordination overhead means one operator captures all the compounding benefits without dilution across a team.

This is why the CEM framework treats AI not as a tool but as an "enabling environment" --- a container that creates conditions where entirely new execution physics become possible. The distinction operates precisely as Goldratt (1984) described in the Theory of Constraints: each removed constraint was a bottleneck limiting throughput. Traditional frameworks optimized within these bottlenecks. AI eliminated the bottlenecks themselves. That is a fundamentally different intervention --- one that invalidates the assumptions underlying every major framework created before 2023.

The CEM portfolio provides the empirical proof point. 596,903 lines of production code. 10 shipped systems. 2,561 commits. 29 commits per active day against an industry median of 2. A 12.1% defect rate against an industry norm of 20--50%. All produced by one operator with no prior development experience, operating in the enabling environment that these five constraint removals created.

What This Means for Engineering Organizations and Solo Operators

For engineering organizations: the frameworks you are using were designed for constraints that may no longer apply to your context. This does not mean Scrum is wrong or that Lean Startup is obsolete --- both remain appropriate where their target constraints still bind. But if your organization is running AI-assisted development inside a framework designed for pre-AI constraints, you are paying coordination overhead, ceremony costs, and planning taxes that no longer buy what they used to buy. The question is not "How do we add AI to our existing process?" The question is "Which of our existing processes exist because of constraints that AI has removed?"

For solo operators and small teams: the constraints that previously required organizations --- expertise breadth, coordination capacity, building capital --- have weakened or dissolved. The operational space available to one person with AI and a structured execution method is larger than it has ever been. The CEM data demonstrates the ceiling: enterprise-scale output from a solo operator in four months. The enabling environment exists. The question is whether your methodology matches the environment you are actually operating in, or the environment that existed before 2023.

---

Related: How to Know When to Kill a Software Project | How Small Development Cycles Build Large Software Systems Faster

References

1. McKinsey & Company (2023). "The Economic Potential of Generative AI." Software development impact analysis and automation estimates.
2. GitHub (2023). "AI in Software Development Survey." Developer workflow changes and productivity patterns with AI assistants.
3. Bain & Company (2023). "AI Engineering Productivity Data." Task-level acceleration patterns across engineering organizations.
4. Carta (2025). "Solo Founders Report." Founder demographics and venture data (solo-founded startups: 17% in 2017 to 36% in 2024).