Thoughtworks Technology Radar, Volume 34 Released

Thoughtworks Technology Radar Volume 34 has been released. This Radar organizes technology trends into four stages—Adopt / Trial / Assess / Hold—and focuses in particular on technology assessment and operational approaches suited to the AI agent era.

The core themes fall into four main areas.

• The agent era and technology assessment: New terms and tools are emerging so quickly that the challenge of assessing them before their meaning solidifies has grown larger. Concepts like spec-driven development and harness engineering have blurred boundaries, and tool lifespans have also shortened, making sustainable judgment difficult.
• Keep the principles, but revisit the patterns: As AI rapidly generates software complexity, fundamental principles like clean code, testability, accessibility, and intentional design are re-emphasized. Pair programming, zero trust, mutation testing, and DORA metrics are also revisited.
• Security issues with agents: The more valuable an agent is, the more permissions it needs—but this also increases the risk of prompt injection and data leaks. The lethal trifecta problem, least privilege, zero trust, and sandboxing are presented as basic premises.
• Coding agent harnesses: Instead of the temptation to remove humans from the loop, mechanisms are needed to guide and verify agent behavior. Agent Skills, spec-driven development, feedback sensors, deterministic checks, and measuring collaboration quality are presented as important pillars.

Notable items introduced under Adopt include Context engineering, curated shared instructions for software teams, DORA metrics, Passkeys, Structured output from LLMs, and Zero trust architecture.

Under Trial, notable items include Agent Skills, browser-based component testing, feedback sensors for coding agents, mutation testing, sandboxed execution for coding agents, semantic layer, and server-driven UI.

Under Assess, topics covered include agentic reinforcement learning environments, architecture drift reduction with LLMs, code intelligence as agentic tooling, context graph, feedback flywheel, HTML Tools, semantic entropy-based LLM evaluation, measuring collaboration quality with coding agents, MITRE ATLAS, and Ralph loop.

Overall, this Radar views AI not merely as a productivity tool, but as a technological shift that forces teams to redesign context management, security, verification, and team operations. The message is strong that precise control and sustainable engineering matter more than rapid generation.