Critical Propulsion
← Back to Insights
DeliveryCritical Propulsion12 min read

Greenfield vs. Brownfield. One Swarm. Two Playbooks.

Whether you're building from scratch or modernising what already exists, AI agent swarms adapt to your reality. But the approach, the risks, and the ROI curve are fundamentally different.

70%
of AI-generated code accepted in greenfield projects vs. 30% in legacy codebases
3-5×
faster initial scaffold generation with agent swarms
60-80%
of enterprise IT budgets go to maintaining existing systems

Profound Logic, industry benchmarks

60-80%
of legacy modernization projects stall or fail to deliver expected ROI

Tricension, McKinsey

Most Enterprises Are Running Both at Once

If you're a CTO or VP Engineering at a mid-to-large enterprise, you're probably managing both: a greenfield initiative (new product, new platform, new capability your board is demanding) and a brownfield reality (the legacy systems that still generate 80% of your revenue and 90% of your headaches).

Most delivery models treat these the same way. Same team shape, same process, same rate card. They're actually different problems with different operating parameters. Greenfield and brownfield projects have fundamentally different risk profiles, different bottlenecks, and different value curves. The AI agent swarm that accelerates one can derail the other if deployed incorrectly.

We run two distinct playbooks, both powered by the same Pulse delivery framework and Agentic Swarm model, tuned for the specific realities of each project type.

Brownfield: Modernize Without Breaking
  • Legacy constraints are real (existing data, APIs, dependencies, users)
  • Agent code acceptance requires deep context understanding first
  • Discovery-heavy Pulse cycles, understand before you build
  • Architecture agents learn existing patterns before enforcing new ones
  • Peak risk: breaking production, data migration, regression, business continuity
  • Swarm advantage: comprehensive testing, parallel workstreams, risk reduction
Greenfield: Build Fast, Build Right
  • No legacy constraints, architecture designed for AI from day one
  • Agent code acceptance rate is 2-3x higher (no legacy patterns to navigate)
  • Full Pulse cycle from discovery through deployment
  • Architecture agents enforce standards before technical debt exists
  • Peak risk: scope creep, over-engineering, building the wrong thing
  • Swarm advantage: speed to first working version, rapid validation

How the Swarm Adapts: Greenfield vs. Brownfield

DimensionGreenfield DeploymentBrownfield Deployment
Week 1 FocusArchitecture design, scaffold generation, CI/CD pipelineCodebase analysis, dependency mapping, risk assessment
Agent Onboarding1-2 days — agents start generating from architecture specs5-7 days typical — agents learn existing patterns, constraints, data models (varies with codebase size)
Domain AgentsCode Gen, Architecture Enforcement, Story Writer, QACode Analysis, Regression Testing, Migration Planning, Dependency Mapping
Code Acceptance Rate60-70% — clean codebase, no legacy patterns25-40% — must navigate existing patterns and constraints
Testing StrategyTDD from day one, agents generate test suites alongside codeCharacterization testing first, then regression, then new feature tests
Biggest RiskBuilding the wrong thing fastBreaking what already works
Pulse Cycle Length5-days — rapid build-validate loops5 days — extra validation and regression gates
Human Oversight FocusArchitecture decisions, product alignment, scope controlRisk assessment, data migration validation, business continuity
Time to First ValueWeek 2-4 — working prototype (typical)Week 6-8 — first modernized component in production (typical)
Effective Capacity MultiplierExpected 3-5x vs. human-only teams2-3x vs. human-only teams (more human judgment required)

Right-Sized Engagement Models

Greenfield: Why AI Swarms Are a Natural Fit

Greenfield projects are where AI agent swarms deliver the most dramatic ROI. No legacy code to navigate. No brittle dependencies to protect. No years of technical debt to work around. The swarm starts with a clean architecture and generates at full velocity from day one.

But greenfield has its own trap: speed without alignment. The number one risk isn't technical. It's building the wrong thing fast. This is why the Pulse framework matters more in greenfield than anywhere else. 5-day cycles of discover-build-validate-learn prevent the classic greenfield failure: a beautiful, well-architected system that nobody asked for.

Week 1-2
Architecture Sprint
Product Engineering Architect + Architecture Agent define the target state. AI generates the scaffold — project structure, CI/CD pipeline, database schemas, API contracts — while the human validates every design decision.
Architect AgentCode Gen Agent
Week 2-4
Foundation Build
Swarm generates core functionality. AI agents write code, tests, and documentation in parallel. Human developers review, refine, and handle the nuanced business logic that requires domain judgment.
Code Gen AgentQA AgentStory Writer
Week 4-8
Feature Velocity
With the foundation stable, the swarm enters full velocity mode. Features ship in 5-day pulses. Architecture agents enforce standards on every PR. QA agents run continuous regression. Human leads focus on stakeholder alignment.
Full Swarm
Week 8+
Scale & Optimize
Delivery metrics guide optimization. The swarm identifies bottlenecks, suggests architectural improvements, and adapts agent configurations based on what's working. Human oversight shifts from building to scaling.
Observability AgentOptimization Agent
With this model:
Working prototype typically by Week 2-4, depending on scope complexity. Production-ready MVP typically by Week 8-10. Traditional consulting: first demo in Week 6-8, production-ready in Week 14-18.

Brownfield: Where AI Swarms Reduce Risk, Not Just Speed

Brownfield is harder. Everyone knows it. The codebase is messy. The documentation is outdated (or nonexistent). Business logic is buried in stored procedures written in 2009 by someone who left in 2012. Dependencies are fragile. Users depend on behavior that was never intentionally designed.

This is where most AI-assisted development falls apart. Copilots and code generators trained on clean patterns produce suggestions that break when applied to legacy code. The agent acceptance rate drops from 60-70% to 25-40% (MSR 2026 Mining Challenge; Netcorp Industry Aggregate, 2025). The naive approach ("just point AI at the old code and modernize") creates more bugs than it fixes.

Our brownfield playbook is different. We don't start by generating code. We start by understanding.

Week 1-2
Discovery & Mapping
Analysis agents map the existing codebase: dependency graphs, API contracts, data flows, business logic patterns. Our architects identify risk zones and migration boundaries. No code changes yet — just understanding.
Analysis AgentDependency Mapper
Week 2-4
Characterization Testing
Before changing anything, QA agents build a comprehensive test suite that captures the system's current behavior — including the bugs that users have learned to work around. This is your safety net.
QA AgentRegression Agent
Week 4-8
Strangler Fig Migration
New components built alongside old ones using the strangler fig pattern. AI agents generate the new code while maintaining full backward compatibility. Each component is validated against characterization tests before cutover.
Code Gen AgentMigration Agent
Week 8+
Progressive Cutover
Components are cut over incrementally — never big-bang. Traffic is shifted gradually. Monitoring agents watch for regression. Rollback is instant. Each cutover is a Pulse cycle with full validation gates.
Release AgentMonitoring Agent
With this model:
Full codebase mapped in Week 2. Characterization test suite by Week 4. First modernized component in production by Week 6. 60-80% of legacy modernization projects fail. Our discovery-first, test-first approach directly attacks the failure modes.

Our Commitment: Risk-Share on Brownfield Modernization

Legacy modernization is risky. We know it. You know it. Here's how we share that risk:

Characterization Test Guarantee. If a regression bug slips past our QA agents into production within 30 days of component cutover, we triage and remediate immediately as a priority escalation.

Milestone-Based Billing. Billing is tied to milestones and delivered output. We are committed to hitting success criteria before moving forward, and we work with you to resolve blockers when they arise.

Progressive Cutover with Instant Rollback. Every component cutover includes automated rollback capability. If the new component underperforms, traffic reverts to the legacy system within minutes. No manual intervention required.

Performance SLA. For performance-sensitive applications, we baseline existing system performance metrics and guarantee the modernized components meet or exceed those baselines before cutover.

Risk Profile: Know What You're Walking Into

The reason one swarm model can't serve both project types is that the risks are completely different. Here's where each project type concentrates its risk, and how the swarm mitigates it:

Risk DimensionGreenfield RiskBrownfield Risk
Scope CreepHigh — no constraints means endless possibilitiesMedium — existing system defines boundaries
Production BreakageLow — no production to breakHigh — every change risks regression
Data MigrationLow — clean data modelHigh — data quality, schema evolution, backfills
Architecture DriftMedium — agents enforce but scope creep causes driftMedium — existing debt plus new patterns coexisting
Knowledge TransferLow — agents hold all context from inceptionHigh — tribal knowledge, undocumented decisions
Business ContinuityLow — no existing usersHigh — real users, real revenue, real consequences
AI Agent EffectivenessLow — clean code, high acceptance rateMedium — legacy patterns reduce AI accuracy

What the Research Actually Shows

AI-augmented development is transforming how software gets built, but the gains aren't automatic. Our approach is grounded in what the research actually shows: organizations that achieve structured AI adoption, strong workflows, and deliberate process design see measurable improvements. Those that simply deploy tools without workflow integration often see no gains at all. Here's what the data supports:

Greenfield: New Platform Builds

When teams start clean and adopt AI with intent, the productivity curve bends fast:

MetricResearch-Backed Outcome
Delivery Velocity1.5–2x traditional pace
Task CompletionUp to 55% faster on defined coding tasks
Developer Output25–55% more output per developer
Time-to-Value35–50% reduction in key development phases
High-Adoption Orgs110%+ productivity gains (80–100% tool adoption)

Sources: GitHub/Microsoft Research, McKinsey, MIT Sloan

Brownfield: Legacy Modernization

Modernization is harder, but the research shows structured AI adoption still moves the needle meaningfully:

MetricResearch-Backed Outcome
Modernization Velocity1.4–1.6x traditional pace
Timeline Acceleration40–50% faster modernization timelines
Tech Debt Reduction~40% reduction in tech-debt-related costs
Migration Tasks30–60% productivity gains with structured workflows
Developer Output20–30% more output per developer

Sources: McKinsey, Altimi Research, Deloitte, IBM

Why Critical Propulsion Instead of Your Current Vendor

Greenfield and brownfield need different operating shapes. Here's how we structure each:

Two Distinct Playbooks. We don't repurpose greenfield processes for brownfield or vice versa. Different risk profiles demand different agent configurations, different testing strategies, and different stakeholder engagement models.

AI-Native, Not AI-Adjacent. Traditional firms bolt AI tools onto existing delivery models. We built our delivery model around AI agents from the ground up. The difference: they get 10-15% productivity lift. We target 2-5x effective capacity.

Output-Based, Not Headcount-Based. We don't bill for butts in seats. We bill for components delivered, milestones met, and measurable output achieved.

Knowledge Persistence. When a delivery team rotates, institutional knowledge moves with them. In our model, knowledge lives in agent configurations that persist across engagement cycles. Capability that doesn't rotate.

What We Don't Do (And Why That Matters)

Knowing what we won't take on matters as much as knowing what we deliver:

  • We don't rewrite your data model in 6 months. That's an 18-24 month journey and we won't pretend otherwise.
  • Staff augmentation isn't where we differentiate. If you're looking for delivery accountability and measurable output, that's where our model shines.
  • We don't promise full legacy replacement. We modernize component by component, proving value at each stage before expanding scope.
ShareLinkedInX

New Build or Legacy Modernization. The Swarm Has a Playbook for Both.

Whether you're launching a new platform or untangling a decade of technical debt, see how Critical Propulsion's AI-augmented delivery model adapts to your reality.