Thoughtworks Radar 33 and 34: From MCP Everywhere to Agents on a Leash

The Thoughtworks Technology Radar is my trigger to act on a tool (see how I stay current). Read together, the last two volumes are a neat before-and-after of a year spent putting agents into real delivery.

Vol. 33 (November 2025): the agent year

All four themes were about AI: GPU orchestration, agents elevated by MCP, how to feed knowledge to coding agents (AGENTS.md, Context7), and a first list of AI antipatterns: shadow IT, naively wrapping every API as an MCP server, text-to-SQL, and complacency with AI-generated code.

Worth noting: pre-commit hooks and curated shared instructions in Adopt, Claude Code entering Trial, context engineering and spec-driven development still in Assess, and Fastify in Adopt for Node back ends.

Vol. 34 (April 2026): retain the principles

The headline is codebase cognitive debt: the growing gap between how much code AI generates and how much of it the team actually understands. The answer isn’t a new tool. It’s the fundamentals: pairing, zero trust, mutation testing, DORA metrics (now with a fifth, rework rate), testability and accessibility.

The four themes:

1. Evaluating tech in an agentic world. Terms diffuse quickly and many tools are weeks old, so judgement gets harder.
2. Retaining principles, relinquishing patterns. The terminal is back as a primary interface, and team design now includes agents.
3. Securing permission-hungry agents. Useful agents want broad access; zero trust, least privilege and sandboxes are the floor, and safe systems are pipelines of constrained agents rather than one all-powerful one.
4. Putting coding agents on a leash. Harnesses: feedforward controls (skills, shared instructions, specs) plus feedback controls (compilers, linters, tests, mutation testing) so an agent corrects itself before a human reviews.

The outer ring is renamed from Hold to Caution.

What moved

• Context engineering: Assess → Adopt.
• Claude Code: Trial → Adopt.
• Structured output from LLMs: Trial → Adopt.
• AG-UI (agent–UI protocol): Assess → Trial.
• LangGraph: Adopt → Trial.
• MCP: from a Trial platform to “MCP by default” in Caution. Use it where a structured tool contract and auth earn their keep; often a CLI, an API or an Agent Skill is simpler.

What I’m taking from it

• Harnesses over heroics. Agent Skills and Harness Engineering is exactly the “leash” theme, and it’s how I run my own second brain: a small constitution file, skills that load when needed, and git after every run.
• Measure outcomes, not output. Coding throughput stays in Caution; DORA with rework rate is the honest alternative.
• Treat agent access as a security boundary. Sandboxed execution moves to Trial, which is the argument in A Safe Path for AI Agents in a Large Codebase.
• Verification is the job now. Cognitive debt is why Don’t Read the Code? Verifying AI-Written Code matters.

My own placements live on the radar. Vol. 35 is due this autumn; I’ll add it when it lands.

Sources: Vol. 34 PDF · Vol. 33 PDF · thoughtworks.com/radar