Software engineering
SWE-bench
Catalyst is evaluated on SWE-bench — resolving real GitHub issues against real repositories — as the public yardstick for end-to-end code work.
Technology / Legacy Modernisation
Forty-year-old code, moved into this decade — without stopping the business.
Catalyst is our proprietary multi-agent system and MCP for migrating and modernising legacy codebases. COBOL and IBM mainframe to Java and modern frameworks — verified, not hoped.
Why legacy modernisation is hard.
Trillions of lines of COBOL still run banking, insurance and government systems. The risk is never the language — it is that the behaviour is undocumented, the experts have left, and the system cannot be switched off while it is being replaced.
Catalyst treats modernisation as an engineering problem with a verification loop — not a translation exercise you check by eye.
Three reasons the rewrite keeps getting postponed.
The code still runs the business
COBOL on a mainframe is not a museum piece. It is processing transactions tonight. You cannot stop it to rewrite it.
The people who wrote it are gone
The original authors retired. The documentation is partial. The behaviour lives only in the source — and in production.
A rewrite is a bet most teams lose
Big-bang rewrites slip, blow budgets, and re-introduce bugs that the old system had quietly fixed twenty years ago.
Catalyst · proprietary multi-agent system & MCP
Specialised agents, coordinated through a shared knowledge layer.
Catalyst is not one model translating files. It is a set of specialised agents — each scoped to a stage, each governed by our internal protocol for agent quality, readiness and security — working over a shared map of the codebase.
01
Dependency mapping
Catalyst reads the whole codebase and reconstructs the call graph, data flows and hidden coupling — the map nobody on the team currently holds in full.
02
Architectural analysis
It recovers the implicit architecture: what the modules really do, where the boundaries are, which behaviours are load-bearing and which are dead.
03
Refactoring & migration planning
A sequenced plan — strangler-fig where it fits, module-by-module where it does not — so the system keeps running while it is modernised.
04
Translation
COBOL, IBM mainframe code and other legacy stacks translated into Java and modern frameworks — preserving behaviour, not just syntax.
05
Test generation
Characterisation tests pinned to the legacy behaviour first, so the migration is verified against what the old system actually did — not against a spec that may not exist.
06
Verification
Outputs validated against the generated tests and against provenance. Every change is traceable to the legacy behaviour it preserves or deliberately changes.
Measured, not asserted.
We hold Catalyst to public and internal benchmarks so a migration claim is backed by evidence a client can inspect.
Internal · legacy translation
Migration benchmarks
Internal task batteries on COBOL-to-Java and mainframe migration: behaviour preservation, test pass-through and review effort measured per module.
Full benchmark methodology and results are being written up in a technical paper — in preparation. Reach out if you want the current figures under NDA.
The research underneath Catalyst.
Catalyst applies the same multi-agent architecture and verification discipline we research across KVA — role-scoped agents, audited workflows, capability gated behind evidence.
Have a legacy system you cannot turn off?
We start with a Catalyst assessment: a dependency map and a sequenced modernisation plan, before a single line is migrated.