Claude Opus 4.8’s Dynamic Workflows shipped with a capability that restructures how enterprise modernization programs can be staffed: hundreds of parallel subagents running inside a single session, capable of executing codebase-scale migrations across hundreds of thousands of lines of code. That’s a meaningful lift in raw throughput.
But throughput was never the constraint in enterprise software modernization. Understanding why changes how you evaluate this capability.
mindmap
root((Modernization<br/>with AI Agents))
AI handles well
Code pattern migration
Test suite generation
Documentation from code
API contract extraction
Humans must own
Architecture sequencing
Data migration timing
Stakeholder cutover decisions
Production validation
What Parallel Subagents Actually Do
Dynamic Workflows lets an AI coding session decompose a large codebase task into parallel workstreams and execute them simultaneously. A modernization that involves migrating a consistent code pattern across thousands of files, generating test coverage for a large undocumented module, or extracting API contracts from legacy integrations — these are tasks where the parallel agent model applies directly.
The throughput gains are real. Work that would take a development team several weeks of coordinated effort can be compressed into hours of supervised AI execution. That’s a genuine change in the economics of the assessment and incremental migration phases of a modernization program.
What it doesn’t change is the architecture and sequencing layer — which is where enterprise modernization programs usually stall.
The Actual Bottleneck in Modernization Programs
The LERETA modernization ran from 2020 through 2024 — a four-year, $20M investment in rebuilding two flagship products for the second-largest property-tax processor in the United States, processing roughly $18 billion annually. At peak, I was leading thirty developers across multiple teams.
The bottleneck in that program was never coding throughput. The team was capable. The bottleneck was sequencing: which system do you touch first, which dependencies have to move together, which legacy data structures have to be preserved until the new system is fully stable, which decisions about the target architecture haven’t been made yet and are blocking everything downstream.
Getting the board to commit to a $20M modernization required presenting wall-sized architecture diagrams that made the full legacy-to-modern critical path visible. That architecture work — the AS-IS documentation, the integration mapping, the dependency graph, the sequencing decision — was the hard part. The coding itself was downstream of those decisions.
Parallel subagents accelerate the coding. They do not resolve the architecture. They do not sequence the data migration. They do not make the cutover decisions that depend on production validation of a new system running in parallel with the one it’s replacing. All of that remains human work.
What This Means for Enterprise Modernization Planning
The practical implication is that AI-assisted modernization shifts the value of a fractional CTO or principal architect engagement toward the front end of the program.
In a traditional modernization, significant engineering hours go into rote migration work — the pattern changes, the framework upgrades, the test generation. Parallel AI agents compress that work substantially. What they don’t compress is the discovery and architecture phase — the work of understanding what you have, deciding what you want, and sequencing the path between them.
That work is now a larger fraction of the total program because the back end compressed. Organizations that plan their AI-assisted modernizations the same way they planned traditional ones will find themselves over-staffed in the migration execution phase and under-resourced in the architecture and validation phases.
The LERETA program also ran into a concrete version of the retrofitting vs. clean-rebuild decision. We acquired a Texas company with similar technology and attempted to retrofit it into the flagship product rebuild rather than designing a new architecture from the start. The data structures differed in subtle ways; the processing logic was fundamentally different. The retrofit created rework and delay that a clean-slate approach would have avoided.
AI-assisted migration doesn’t resolve the retrofitting vs. rebuild tradeoff. It makes the execution of either path faster. The decision about which path to take remains architectural, and getting it wrong is expensive regardless of how quickly you can execute the wrong path.
What to Do With This Capability Now
If you’re running or planning a modernization program, the Dynamic Workflows capability changes three things practically.
The assessment phase can go deeper. AI agents can extract integration contracts, generate dependency maps, and produce documentation from undocumented legacy code at a speed that was previously impractical. An organization that previously couldn’t justify a full code-level assessment of a complex legacy system can now commission one at reasonable cost. That changes the quality of the information going into the architecture decisions.
The incremental migration phase runs faster. Pattern-based migrations that would have required months of coordinated developer effort can be executed in parallel with AI supervision. That compresses the program timeline on the execution side.
But the architecture, sequencing, and validation work expands in relative terms. That’s where human oversight is irreplaceable, and where the decisions that determine whether a modernization program delivers its intended outcome get made. Organizations that treat Dynamic Workflows as a way to reduce the architecture and planning investment are likely to discover the same problem LERETA avoided — moving fast on a path that hadn’t been sufficiently thought through.
The Architecture Remains Human Work
The Dynamic Workflows announcement represents a genuine capability threshold — the kind of shift that changes the economics of modernization programs. The organizations that capture the most value from it will use it to do more thorough discovery and faster execution while maintaining human ownership of the architecture and sequencing decisions.
The organizations that miss the value will use it to execute faster on plans that haven’t been sufficiently thought through. Fast execution of the wrong architecture is still the wrong architecture.
