Agentic Software Development
When it has to work.
Agentic coding holds up only when it operates against a deterministic model of the system. We re-develop legacy applications with agentic coding, grounded in a trustworthy model, so the result holds up against systems you cannot afford to get wrong.
Grounded, not guessing
Point agents at an undocumented system and they hallucinate, and a confident wrong answer against a core system is expensive. We give them a deterministic, source-cited model to reason against. Through Sysparency MCP server the agent pulls in the business context itself: what a program is for, where it sits in module and process, who depends on it. That is how the agent plans changes in their business context, and the result can be checked.
What the agent gets through Sysparency MCP server
- Business context: the module, the process step and what the program is for, not just what it does.
- Every dependency: callers, tables written, interfaces on the same data.
- Real usage: how often, by whom, dialog or batch.
- Evidence: every fact cited to its source line, so every change can be checked.
What we deliver
Re-engineer legacy with agentic coding
Re-develop a legacy application against a deterministic model of its real behaviour — preserving the rules that matter, minimising the risk.
Agentic AI across the SDLC
Embed agentic workflows into specification, generation, verification and delivery — with humans in control of critical change.
Stand up your agentic pipeline
Set up the agentic software development pipeline inside your organisation, on your systems and standards.
Example
25 year old separate transactions. One launchpad. And a better system.
Twelve SAP GUI module pools of a legacy application were turned into a launchpad with three Fiori Elements apps, a dashboard and four RAP business objects. Fully agentic: the agent understood the legacy programs through Sysparency MCP server and built the new system through SAP ADT MCP server; not a single line was written by hand. The graph decided what was built differently: rules once in the data model instead of in twelve programs, one click path instead of three transactions, one repair run fewer, one dead field fewer, four analytics that never existed.
Left the transaction list of the legacy programs, right the launchpad that replaces them, built end to end by an agent without hand-written code. Legacy application on a trial system.
- 12 → 4
- programs → business objects, one OData service
- 3 → 1
- transactions per booking → one click path
- 0 → 7
- rules on save, instead of a nightly check run
- 0 → 4
- analytics: KPIs, charts, ranking, trend
| What the whole picture showed | What it became | Evidence from the graph |
|---|---|---|
| The dialogs are not the only writers to the booking table. | The business objects become the single entry with the rules; BAPI, ALE and IDoc are the next step, the list already exists. | Writers from the graph: dialog, function group SAPBC_GLOBAL_FIS, BAPI group, ALE generator; 35 RFC-enabled modules on the same tables. |
| A nightly check run repairs what the entry point fails to validate. | Existence checks for all three master objects fire on save; the check run is retired. | Check program with usage 0 and lookups on the three master tables; ATC priority 1 on the edit dialog. |
| 73 % of all calls are display only. | Read mode as default, editing as a deliberate step; a retirement list with usage evidence, the obsolete smoker flag removed. | Usage counters per transaction from the graph: 4,870 display calls versus 158 agency edits. |
Built on SAP reference data on a trial system, not on a customer system. RFC, ALE and IDoc still write to the tables directly; that is the documented next step.
Request the full migration reportFor AI platform teams
Sysparency understands the as-is. SAP's tools write. Your agent connects the two.
The Sysparency MCP server exposes the knowledge graph of your custom code to any MCP-capable client: Claude, ChatGPT, Copilot or your own framework. The agent reads facts with evidence instead of guessing at ABAP, and hands over to SAP's ADT tools, which generate, activate and transport. Add SAP's documentation MCPs and the agent looks up the standard for every extension: what the standard offers today, what a Z program merely replicates, where a real gap remains. Back to standard and gap analysis, with evidence on both sides.
Seven read-only tools
- Graph schema
- Cypher query
- Semantic search over business descriptions
- Object list
- Metadata
- Read source
- Search source
Read only, your code only
Every tool is read-only. The graph holds only your custom code (Z, Y, your namespaces), never SAP standard.
A dedicated environment per customer
Graph and server run in a dedicated environment, access through Microsoft Entra ID, EU hosting or your own infrastructure.
Built into the report
The connection dialog sits in the report header: copy the URL, connect your client, ask.
Back to standard, gap analysis
Sysparency MCP server knows the as-is of your custom code, SAP's documentation MCPs know the standard. Together the agent finds the standard replacement for an extension and documents the gap that remains.
Across systems
One agent, several SAP systems.
Every system gets its own knowledge graph and its own MCP server. Connect two or more of them to one agent and it works across systems: it compares extensions, finds what exists twice, shows where copies drifted apart and what a template rollout or a merge needs. Every finding is cited from both graphs.
- Compare: the same Z object in two systems, differences down to the line.
- Consolidate: which extensions can become one template, which stay local.
- Roll out: what a target system is missing before a template can land.
Questions about agentic re-engineering
Straight answers.
What platform and architecture teams ask before their first agentic rebuild.
An AI agent understands the legacy programs through Sysparency MCP server, decides from the graph what to build differently, and builds the new system through SAP ADT MCP server. No line is written by hand.
Without it the agent sees one file and guesses the rest. With the graph it knows every writer of a table, every caller, the real usage and the business meaning, each cited to the source line.
Any MCP-capable client: Claude, ChatGPT, Copilot or your own framework. The seven tools are read-only and only cover your own code.
Yes. With SAP's documentation MCPs connected, the agent looks up what the standard offers today, what a Z program merely replicates and where a real gap remains.
It is a legacy application on an SAP trial system, not a customer system. Twelve module pools became one launchpad with three Fiori Elements apps, a dashboard and four RAP business objects.