ThoughtWorks Radar highlights multi-agent orchestration and harness engineering

ThoughtWorks Technology Radar Volume 35 identifies four primary themes, all related to LLMs and AI agents. The first theme, 'Putting More Agents to Work,' addresses the shift from single-agent CLI usage to managing multiple parallel coding agents. This is driven by model capabilities that allow longer autonomous tasks, reducing the need for turn-by-turn human oversight. New tooling like Claude Squad, T3 Code, and Orca provides meta-harnesses to visualize agent status and manage workspaces, such as Git worktrees.

The second theme, 'Harness Engineering at Scale,' focuses on the infrastructure surrounding agents, defined as everything except the model, including skills, plugins, and MCP servers. A new open-source standard called 'Agent Plugins' has been adopted by major labs to standardize the sharing of MCP servers and agent skills, addressing the previous lack of interoperability between different coding agent harnesses. ThoughtWorks advises against 'human in the loop theater,' where developers blindly approve agent actions due to fatigue, and recommends using sandboxing, least privilege, and automated feedback loops to maintain quality without constant manual intervention. The report argues that developers should focus on improving the agent harness and automated feedback loops rather than simply increasing the number of parallel agents, and that organizations must treat harness engineering as a platform discipline to avoid redundant work and ensure accountability.