Thoughtworks Technology Radar Vol.34: Agentic Coding Matures and React Returns to Adopt

TLDR: Thoughtworks published volume 34 of its Technology Radar, and the recurring theme is that agentic software engineering is moving from experiment to discipline. React JS and React Native both jumped to Adopt after a decade away, lakehouse architectures keep growing, and AI security has become its own category of concern.

Summary: The framing of this volume is honest in a way I appreciate. AI dominates and is bleeding into every other area of software, but the authors push back gently against the idea that AI is the only story worth telling. The headline in the agentic engineering section is what they call harness engineering, which is the practice of building reliability and consistency around coding agents rather than just trusting them to behave. Several entries sit in Trial that point at this: agent skills, which are packaged scripts and resources that let an agent complete a task more reliably; feedback sensors for coding agents, which act as deterministic quality gates that trigger an agent to self-correct; and sandboxed execution, isolating agents so a bad run cannot damage the system or leak credentials. That last one should be table stakes, and I am glad to see it called out explicitly.

What graduated to Adopt is telling. Claude Code and Cursor are both there, alongside context engineering, which they treat as a core part of building an agent harness. The cautions are where the real engineering judgment lives. They flag agent instruction bloat, where you stuff so much into the prompt that the agent gets worse, not better. They flag ignoring durability in agent workflows, which bites you the moment an agent is running a long multi-step task and something fails halfway. And they put using MCP by default in the caution ring, which is a sharp call given how fashionable MCP has become. Reaching for a protocol just because it exists is not architecture, and someone needed to say that out loud.

Two cultural cautions go deeper than tooling. Measuring productivity by coding throughput and codebase cognitive debt both speak to the same trap: in an AI era you can generate enormous volumes of code quickly and convince yourself that means progress. The Radar revisits DORA metrics in Adopt for the fourth year running, now reframed around AI-readiness, which is a direct rebuttal to throughput thinking. There is something the author is dancing around here, though. If agents make it trivial to produce code, the bottleneck shifts entirely to review, comprehension, and the slow human work of understanding a system. The Radar names cognitive debt but does not really wrestle with who pays it down or how.

The data and analytics section is shaped almost entirely by what AI systems need. The semantic layer sits in Trial, and its job is to organize and label data so it is legible to agents rather than just to humans. PageIndex and Graphiti address the agent-data relationship, with Graphiti being a temporal knowledge graph that tracks how facts change over time to keep LLM responses grounded. Lakehouse architecture continues its march, with Apache Iceberg in Adopt for vendor-neutral lakehouses and DuckLake in Assess as a lower-complexity alternative. On cloud and infrastructure, the cost story is loud: OptScale for tracking and optimizing GPU and ML spend, RunPod as a distributed GPU marketplace, and SigNoz as an open-source, OpenTelemetry-native alternative to vendor monitoring. mise lands in Adopt as a Rust-based tool that consolidates environment variables, language versioning, and task execution, which fits the broader drift back toward the terminal that AI has accelerated.

The app and web development section is the one that made me sit up. React JS enters Adopt, driven largely by positive experience with React Compiler since its October 2025 release, and React Native enters alongside it. Both last appeared on the Radar ten years ago. That is a genuinely interesting signal about maturity cycles. A framework can be ubiquitous and still fall off a curated Radar for a decade, then return because the surrounding ecosystem finally caught up to its promise. Svelte also sits in Adopt as a lighter alternative with production success, and TanStack gets a nod. On security, the rapid adoption of agents brings real risk: OpenClaw is flagged in Caution as a prime example of a permission-hungry agent, MITRE ATLAS is in Assess as a resource for spotting agentic threats, and toxic flow analysis sits in Adopt.

Key takeaways:

Harness engineering is the organizing idea for agentic coding: build reliability around agents with skills, feedback sensors, and sandboxed execution rather than trusting raw output.
Using MCP by default is now a caution, a useful reminder that adopting a protocol is not the same as solving a problem.
React JS and React Native returned to Adopt after ten years, with React Compiler cited as the main reason for React's inclusion.
Throughput and cognitive debt are the cultural risks of the AI era; DORA metrics return in Adopt, reframed around AI-readiness.
AI security is now a first-class concern, with permission-hungry agents and toxic flow analysis getting explicit attention.

Why do I care: If you build frontends or own architecture decisions, two things here matter directly. React's return to Adopt with React Compiler as the justification confirms that the compiler is production-credible, which changes how you reason about memoization and re-renders going forward, so it is worth a real evaluation rather than a wait-and-see. The harness engineering theme is the bigger one. As your team leans on Claude Code or Cursor, the differentiator stops being whether you use agents and becomes how disciplined your harness is: sandboxing, feedback gates, lean instructions, and durable workflows. The caution on MCP by default and the warning about measuring throughput are both arguments you will need to make to managers who see agent output volume and assume velocity. This is a strategy and architecture story more than a how-to, but those are exactly the conversations a senior engineer ends up owning.

Thoughtworks Technology Radar Vol.34