Modernisation of legacy systems has historically been constrained by three factors: incomplete documentation, inaccessible source code, and shrinking institutional knowledge. Traditional delivery models treat reverse-engineering as a manual, time-intensive discovery phase — often consuming more budget than the rebuild itself.
Recent advances in agentic AI workflows and multimodal large language models introduce a different operating model: one where legacy estates can be analysed, specified, and re-platformed through automated interpretation of code, interfaces, and human expertise.
This article outlines practical patterns for migrating legacy platforms using AI agents — especially in environments where artefacts are partial, opaque, or operationally entangled.
1. Agent-Driven Legacy Code Analysis
Where source code is available, AI agents can perform structured analysis across multiple layers simultaneously.
Static codebase interpretation
Agents can:
- Parse monolithic repositories
- Identify bounded contexts
- Extract domain models
- Map service dependencies
- Detect dead code and unused modules
This enables automated production of:
- Functional specifications
- Entity diagrams
- API contracts
- Data flow maps
Rather than reading code linearly, agents reason across the system holistically — clustering related logic and inferring business purpose.
Database schema reconstruction
Legacy logic often resides in:
- Stored procedures
- Triggers
- Views
- Batch jobs
Agents can analyse DDL + procedural SQL to derive:
- Business rules
- Validation logic
- Workflow triggers
- Data lineage
This is particularly valuable in Oracle, DB2, or SQL Server estates where application layers are thin and behaviour is database-centric.
2. UI Reverse Engineering via Screenshot Analysis
In many legacy environments, source code is incomplete, proprietary, or low-code generated. In these cases, the user interface becomes the primary observable artefact.
Multimodal LLMs allow screenshots to be converted into structured system knowledge.
Screenshot → textual system description
By analysing UI captures, agents can extract:
- Field names and types
- Input constraints
- Dropdown value semantics
- Navigation hierarchies
- Action buttons and workflows
Example outputs:
- Screen inventories
- CRUD operation mapping
- User journey diagrams
- Role-based UI variants
This is especially effective in platforms such as:
- Oracle APEX
- Lotus Notes
- PowerBuilder
- Access applications
- Proprietary desktop clients
Workflow inference
Sequences of screenshots or recordings enable agents to infer:
- Process lifecycles
- Approval chains
- Escalation paths
- Exception handling
This transforms visual interaction into executable system specifications.
3. Audio Capture of SME Knowledge
One of the highest-risk aspects of legacy migration is knowledge attrition — particularly when original system builders are unavailable or knowledge is scattered.
AI enables structured capture of tacit knowledge through recorded dialogue.
SME interview ingestion
Subject Matter Experts can provide verbal walkthroughs covering:
- Business rules
- Edge cases
- Operational workarounds
- Historical design decisions
- Regulatory constraints
Audio is transcribed using speech-to-text pipelines, then analysed by LLM agents.
From conversation to specification
Agents can convert raw transcripts into:
- Functional requirements
- Decision trees
- Process documentation
- Data handling rules
They can also reconcile SME explanations against code or UI artefacts, identifying:
- Undocumented behaviours
- Policy deviations
- Operational risks
This closes the institutional knowledge gap without requiring formal documentation exercises.
4. Multimodal Synthesis: Building a Unified System Model
The real leverage emerges when agents correlate multiple artefact types.
Inputs may include:
- Source code
- Database schemas
- UI screenshots
- Screen recordings
- SME audio interviews
- Legacy documentation
Agents can synthesise these into a unified model comprising:
- Domain entities
- Service boundaries
- Workflow orchestration
- Integration contracts
- Non-functional constraints
This becomes the foundation for re-platforming.
5. Automated Specification Generation
Once analysed, agents can produce structured artefacts required for modern delivery pipelines:
- OpenAPI specifications
- Event schemas
- Entity-relationship diagrams
- Infrastructure requirements
- Security classification models
Specifications can be exported into formats consumable by engineering teams or downstream AI builders.
This eliminates months of traditional discovery workshops.
6. Code Generation and Re-Platforming
With specifications in place, agent workflows can assist in rebuilding target systems.
Typical transformations include:
| Legacy Source | Target Platform |
|---|---|
| PL/SQL logic | Java / Spring services |
| Forms UI | React / Angular |
| Batch jobs | Event-driven workers |
| Monolith DB | Service-scoped schemas |
Agents can generate:
- Service scaffolding
- Data access layers
- Validation rules
- Migration scripts
Human engineers remain responsible for architecture and quality control, but delivery velocity increases materially.
7. Test Case Derivation
Legacy systems often lack formal test coverage.
Agents can derive tests from:
- Validation logic
- Historical defects
- UI constraints
- SME narratives
Outputs include:
- Unit test suites
- Integration scenarios
- Regression packs
- Synthetic data sets
This ensures behavioural parity between legacy and modernised platforms.
8. Documentation Regeneration
Documentation gaps are endemic in legacy estates.
AI agents can automatically produce:
- System overviews
- Runbooks
- Support guides
- API usage documentation
- Operational procedures
This shifts documentation from an afterthought to a generated asset of the reverse-engineering process.
9. Additional High-Value Use Cases
Beyond direct migration, multimodal legacy analysis enables:
Compliance auditing
Agents can detect:
- Data retention violations
- Access control weaknesses
- Encryption gaps
Licensing risk analysis
Identifying dependencies on:
- Oracle
- IBM
- Proprietary runtimes
Supporting cost-reduction strategies.
Performance bottleneck discovery
Analysing:
- Query inefficiencies
- Batch processing delays
- Concurrency conflicts
Integration mapping
Documenting undocumented interfaces between:
- Internal systems
- Third-party vendors
- Government platforms
10. Delivery Model Transformation
Traditional legacy replacement programmes follow a sequence:
- Discovery
- Documentation
- Design
- Build
- Test
Agentic workflows compress phases 1–3 into parallel, automated processes.
This yields:
- Shorter mobilisation timelines
- Reduced consultancy cost
- Faster architectural convergence
In some cases, single-team or even single-engineer rebuild spikes become viable where large programmes were previously assumed necessary.
Conclusion
Legacy migration is no longer constrained to code translation. With multimodal AI agents, organisations can reverse-engineer entire systems from fragmented artefacts — combining code analysis, UI interpretation, and institutional knowledge capture.
Screenshots become specifications. Conversations become requirements. Database triggers become domain rules.
The result is a materially accelerated path from opaque legacy estates to modern, maintainable platforms — with reduced delivery cost and lower operational risk.
As agent capabilities mature, this approach is likely to evolve from experimental practice into a standard operating model for public sector and enterprise re-platforming initiatives.