AI Radar: Revisit Foundations, Secure Agents, Review Code
Thoughtworks' 34th Radar shows AI dominating tech trends, forcing revisits to core practices like pair programming and clean code to counter generated complexity, while emphasizing security for permission-hungry agents and human review of AI code.
AI Forces Return to Software Foundations
AI tools accelerate complexity generation, prompting developers to revisit established practices as a counterweight. Thoughtworks' 34th Technology Radar, with 118 blips, highlights this: pair programming, zero trust architecture, mutation testing, DORA metrics, clean code, deliberate design, testability, and accessibility regain focus. Command line interfaces resurge as agentic tools make terminals primary for developers, reversing years of abstraction for usability.
Secure permission-hungry agents by addressing prompt injection risks, where models fail to distinguish trusted from untrusted inputs despite broad access needs for tasks like code swarms or real-work supervision. Use harness engineering—guides and sensors—to constrain agents safely; expect more such blips in six months.
Human Oversight Essential for Durable AI Code
AI-generated code like Claude's can produce working Python (unit tests pass, handles complex infra) but balloons files to 50KB (2,000 lines) in 100KB total, leading to hacks like sed edits. Even 500,000 lines from Claude Code leak mixes good architecture with mess—humans must read to discern.
Framework for review: Throwaway analysis scripts tolerate AI slop; maintained tooling and durable code demand regular human checks, even via model evaluation with good-code hints. AI responds to discomfort prompts (e.g., "file too big") by decomposing sensibly, adding classes/tests—but won't volunteer. Use CLAUDE.md seriously and patterns to reduce friction, avoiding frustration loops.
Broader AI and Organizational Lessons
LLMs enable ghostwriting, raising philosophy questions on authenticity. In government tech, DirectFile's death under DOGE reveals reform paradox: simple changes hide deceptive complexity, blocked by incumbents. Public service ethos drives better outcomes than disinterest in users.
IRS woes—25% staff loss, 40% budget below 2010 levels—weaken enforcement; boosting funding more than pays via revenue gains, as historical tax efficiency decided empires (Britain vs. France). Lessons apply to large org tech initiatives.