rshetty

Thoughtworks recently launched a new agentic development platform called AI/works™. The marketing claims are quite bold, specifically targeting the "legacy modernization" problem space (which we all know is usually a nightmare).

According to the announcement and their technical guide, the workflow is:

Ingestion: "Blackbox" reverse-engineering of legacy binaries/code (even without full source access in some cases).

Specification: It generates a "SuperSpec" — a machine-readable functional specification enriched with regulatory/security context.

Forward Engineering: Agents use the Spec to generate new code, tests, and pipelines.

Lifecycle: It claims to support a "3-3-3 delivery model" (Idea to Production in 90 days) and includes self-healing/regenerative capabilities for maintenance.

This sounds like the "Holy Grail" of software engineering, but I am skeptical about how well this works on actual enterprise spaghetti code versus carefully curated demos. "Reverse engineering into a perfect spec" is historically where these tools fail.

I’m looking for insights from anyone who has piloted this or works at TW:

How does the "Code-to-Spec" reverse engineering actually handle heavy technical debt or undocumented business logic?

Is the "SuperSpec" truly editable/maintainable by humans, or does it become a new black box?

How much "human in the loop" is actually required for the 3-3-3 model?

Is this built on public LLMs (Claude/GPT-4) or proprietary models trained on legacy patterns (like COBOL/Mainframe data via their Mechanical Orchard partnership)?

Any details on the reality behind the marketing would be appreciated.

This is an amazing piece of software and thanks for creating this. We heavily use Tile38 for all our Geospatial needs in production.