Thoughtworks convened a multi-day retreat in February 2026, gathering senior engineering practitioners from major technology companies to examine how AI agents are changing software development roles and practices. Conducted under Chatham House Rule, the retreat produced a white paper synthesizing more than twenty breakout sessions — not as a consensus roadmap, but, as the report frames it, "a map of fault lines where current practices are breaking and new ones are forming." The central question posed throughout every session — if AI handles the code, where does the engineering actually go — produced no single answer, but the discussions settled into five near-term themes and several longer-horizon signals.
The most actionable near-term finding is that engineering rigor does not disappear when AI writes code; it migrates upstream. Teams adopting <a href="/news/2026-03-14-spec-driven-verification-claude-code-agents">spec-driven development</a> are rediscovering structured formats like EARS (Easy Approach to Requirements Syntax), state machines, and decision tables because traditional user stories lack the precision AI agents need to produce correct implementations. Test-driven development was singled out as producing dramatically better outcomes from AI coding agents, with the mechanism being specific: writing tests before code prevents agents from generating tests that merely confirm their own broken output. The retreat also identified a nascent "middle loop" of supervisory engineering work forming between the inner loop of coding and the outer loop of delivery — a category of work that, as the report notes, has no established name or tooling yet. Security emerged as a critical gap, with the paper warning that granting an agent email access alone can enable a full account takeover, illustrating how far agent security practices lag behind agent capabilities.
Looking further out, the retreat found that Conway's Law now applies to agent topologies. Organizations that fail to account for agent mobility, specialization, and drift risk having their agent graphs mirror dysfunctional org-chart boundaries, introducing the same communication bottlenecks that Conway's Law originally described for human teams. Knowledge graphs and semantic layers — technologies largely set aside for years — are being rediscovered as the grounding infrastructure needed to make agents genuinely domain-aware. On roles, the retreat described PM, developer, and designer responsibilities as converging, with staff engineers facing elevated system-level judgment expectations and junior engineers described as "more valuable than ever" — a counterintuitive conclusion that generated significant debate both in the room and in subsequent Hacker News commentary, where critics questioned whether organizations would actually blend roles rather than simply ship more features at higher velocity. Self-healing autonomous systems were placed on a two-to-five year horizon, blocked by what the paper terms the "latent knowledge problem" — the widening gap between system complexity and human or agent comprehension, which the retreat labels cognitive debt.
That framing is, in the end, the retreat's clearest answer to its own opening question. Engineering doesn't disappear; it moves to wherever the <a href="/news/2026-03-15-comprehension-debt-the-hidden-cost-of-ai-generated-code">cognitive debt</a> is highest — upstream into specs today, into agent supervision tomorrow, and eventually into whatever comes after both.