Wes Reisz, Technical Principal at Thoughtworks, used a roughly 50‑minute presentation at QCon AI to argue that AI‑first software delivery must be pragmatic: teams should integrate large language models into the standard development lifecycle rather than treating them as afterthoughts. He framed the question as one of engineering discipline — how to gain the productivity benefits of coding agents without sacrificing maintainability or verifiability in production systems.
Reisz illustrated his approach with a client engagement he began about three months earlier: a roughly 16‑person team working for a large U.S. state to build a knowledge graph that ingests rules and regulations. The project layers a deep research agent to surface regulatory content for modern AI‑driven user interfaces, runs inside the client’s existing infrastructure, and has moved AI left across design, coding and verification. Reisz said the team’s tooling stack pairs Claude Sonnet 4.5 with Cursor to support those flows.
A recurring question Reisz addressed was why his team does not deploy a full, unsupervised multi‑agent architecture by default. To answer it he proposed a tactical two‑by‑two decision model whose axes are code longevity and level of automated verification. The model guides practitioners toward supervised or unsupervised agent strategies based on how long code must be maintained and how exacting verification needs are, rather than on hype or novelty.
From that model Reisz distilled a practical rule: align agent strategy to risk and maintenance profiles. Work that demands long‑lived code or strict verification favors supervised agent workflows with tighter human oversight; efforts with shorter lifespans and strong automated verification pipelines can tolerate more unsupervised or multi‑agent experimentation. He emphasized structured collaboration patterns that keep engineering rigor even as automation increases.
To operationalize those patterns, Reisz presented RIPER‑5 — Research, Innovate, Plan, Execute, Review — as a lightweight framework for LLM partnership. RIPER‑5 assigns LLMs specific roles at stages of delivery rather than granting end‑to‑end autonomy, with the stated goal of amplifying engineering discipline, reducing brittle automation and producing reproducible outcomes when coding agents are integrated into standard SDLC practices. The practical implication, Reisz suggested, is that client‑facing engineering teams gain a repeatable decision path: use supervised, verifiable agents where permanence and safety matter, and reserve unsupervised experimentation for lower‑risk, well‑tested contexts. That guidance aims to help teams adopt agentic workflows without undermining long‑term maintainability or verification requirements.
Sources
Replies (0)
No replies in this topic yet.
