Thoughtworks Technology Podcast · Thoughtworks

Inside AI/works™: An agentic development platform

·40 min·2 clips

Shodan compares giving an LLM a huge codebase to an open book exam where you don't know where the answer is.

1. The Thoughtworks Technology Podcast episode 'Inside AI/works™: An agentic development platform' examines Thoughtworks' recently launched AI/works platform, drawing on the firm's client work and roughly 20 years of delivery history. 2. Ricky Zachary, practice area lead for engineering platforms with about five years at Thoughtworks, hosts his first episode after an introduction from Ken, with Bharani Subramaniam (CTO of India and Middle East, 16 years at the firm) and Shodan (head of modernization services, about 18 years) as guests. 3. The core thesis is that generative AI alone is not the answer, and that 'sound engineering around generative AI' is needed to make it 'less hallucinative' and 'more reliable'. 4. Bharani traces the platform's origin to a client in India where 30-40 Thoughtworkers built faster with AI, but work stalled in one lane waiting on steps like opening a port. He describes the platform's three parts: an AI gateway with guardrails, central observability, and governance covering spend limits and lineage of who used which model. 5. Shodan's modernization thesis is that legacy systems have stale documentation, absent SMEs and no automated tests, but the code itself is 'a goldmine' of structured data and the most accurate reflection of production. He says the work is a platform-enabled service rather than a button that outputs a modernized system. 6. Tooling projects the ingested code into a capability map for build-versus-buy decisions and incremental planning, plus an event-storm map built from a knowledge graph. Shodan says the capability map replaces an exercise that previously took weeks with SMEs locked in a room. 7. Bharani says the differentiator is a 'context library', or 'hive mind', of how Thoughtworks teams worldwide solved similar problems. He adds that the open challenges are how to create it, curate it and judge its relevance for a given client problem. 8. The hosts agree 'spec' is an overloaded term, since it merges what a system should do with how it is implemented, and 'there's not a standard specification for the specification'. On ingestion, Shodan says tools handle about 10,000 lines but not millions, and compares LLM context to an 'open book exam' where unknown answer locations make the book useless. 9. On metrics, Bharani advises keeping existing baselines such as DORA or flow metrics before and after AI, and Shodan says compound interest shows that steady improvement matters more than a 60% or 70% gain target. Bharani also proposes interdisciplinary pairing, such as product owner with developer or QA with product owner, as a 'paved path' on the platform. 10. Suited to platform engineers, modernization leads and engineering leaders evaluating enterprise AI adoption.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Thoughtworks Technology Podcast