AI Architecture Updates: April 22, 2026

1. Fowler: Thoughtworks’ 34th Radar Centers on Harness Engineering and Permission-Hungry Agents

Martin Fowler. In his April 21 Fragments post, Fowler draws out two threads from Thoughtworks’ 34th Technology Radar: harness engineering as a distinct discipline — the guides, sensors, and review surfaces around AI agents — and the security exposure of “permission hungry” agents that need broad access to be useful but can’t reliably tell trusted instructions from untrusted input. Fowler notes that the radar is stacking multiple blips in the harness category and expects it to grow, while Jim Gumbley’s addition to the radar team brings direct LLM-security analysis. The piece also surfaces a second-order effect: teams are revisiting long-settled practices like pair programming, clean code, and testing with AI pair-programmers in mind. Source

2. Alpha Signal Dissects Harness Engineering Across OpenAI, Anthropic, and Thoughtworks

Alpha Signal. A companion piece from Alpha Signal treats harness engineering as the next discipline to budget for, arguing that “everything around the model is the harness” and showing how different harnesses produce very different end-to-end scores on the same base model. The reported example — an agent going from 52.8% to 66.5% on a task purely through harness changes — quantifies what practitioners have been saying anecdotally: scaffolding, retrieval, tool schemas, and review loops dominate the delta between a demo and a production-grade agent. Source