Layered, opinionated, designed for safe AI.
Generic AI-and-COBOL pitches usually fail for the same handful of reasons: the AI cannot see inside the programs, it has no memory of prior lessons, it has nowhere to truly run what it writes, and there are no guardrails around the parts of the system that matter most. The Filix methodology is engineered to remove each of those failure modes — not with a deck, but with a real platform behind it.
Authoritative metadata
The metadata server gives the AI a structured view of the legacy estate, so it can quote facts about your code rather than guess at them.
Open layerSpecialised AI roles
A small set of agents, each with a clear, separate job. Routing is deterministic; coordination is explicit.
Open layerLoaded expertise
Domain skills the AI uses on demand, embedding the kind of judgement a senior practitioner would otherwise have to provide in person.
Open layerHard rules
Guardrails that make the safe choice the easy default — and that flag risky operations before they happen.
Open layerCompounding memory
A durable record of decisions and lessons. Each engagement starts further along than the last.
Open layerWhy generic AI + COBOL falls short.
- The AI cannot see inside the programs. Public language models have never seen your structures, your control records, or your client-specific layouts. Without authoritative metadata they invent.
- The AI has no memory of prior lessons. Every session starts from zero. Hard-won discoveries get rediscovered, slowly, every time.
- The AI has no way to truly execute what it writes. Without a real runtime under real data, generated tests can pass against a fake while the business breaks.
- The AI has no guardrails around what matters most. A well-intentioned refactor of the wrong code can quietly destroy work that took days to build.
The Filix methodology composes five layers that close each of those gaps.
The methodology, run as a pipeline.
The same five layers underpin a deterministic delivery loop — the Epic Loop — that lets the platform take on real work without giving an unconstrained model the keys.
The Epic Loop is where the roles and the rules meet. A deterministic controller drives an AI “maker” to build a program’s full, Docker-validated test suite end to end. The controller decides what runs and in what order; the AI does the work inside those rails. Autonomy is graded — from human-gated review up to auto-merge — but a hard denylist means protected or client-owned code is never merged automatically, whatever the setting. Every decision is deterministic, logged and auditable, so autonomy never means opacity. It is a reusable delivery asset, not a one-off script: a compounding capability of the platform itself.
A focused walk through the methodology, on your terms.
A working session covering the platform, the methodology and how they would map to the systems you actually have to modernise.