Thoughtworks Tech Radar Vol. 34: What Actually Matters
The Hold ring is gone, replaced by Caution. Context engineering moved to Adopt, LangGraph moved out, and MCP-by-default is now something to think twice about. Here's what the April 2026 radar says about putting coding agents on a leash.
The April 2026 radar just came out. The "Hold" ring is gone, replaced by "Caution." The whole document tackles one question: how do you evaluate technology when AI evolves faster than you can assess it?
Context Engineering Is Now Foundational
Context engineering moved to Adopt, not just as an optimization trick, but as an architectural concern. Teams are shifting from cramming everything into large context windows to using progressive context disclosure: starting with a lightweight index and pulling in only what's relevant.
Three areas are developing quickly: prompt caching for static instructions, dynamic retrieval loading only necessary MCP servers, and context graphs modeling institutional reasoning as queryable data. Treating AI context as a static text box leads to hallucinations.
Putting Agents on a Leash
The idea of a "coding agent harness" runs throughout the document.
Feedforward controls include: Agent Skills (modular, just-in-time instructions), spec-driven development frameworks, and curated shared instructions tied to service templates.
Feedback controls involve: compilers, linters, and test suites integrated into agent workflows. These trigger auto-correction before human review. Tools like cargo-mutants, WuppieFuzz, and CodeScene fit here.
The feedback flywheel connects everything — it's basically retrospectives for your coding agent setup.
Adopt Now
Claude Code — Used daily in production delivery. CLI agent benchmark. Pair it with curated instructions and rigorous review.
Cursor — Default choice alongside Claude Code. Many developers prefer supervising agents within an IDE.
Passkeys — 15 billion eligible accounts globally. NIST classifies synced passkeys as AAL2-compliant. Avoid SMS OTP fallbacks.
Zero trust architecture — Essential for agent deployments. Practice least privilege, continuous monitoring, and use SPIFFE for identity.
DORA metrics — If lead times don't decrease, faster code generation doesn't guarantee better results. Keep an eye on rework rate.
Apache Iceberg — Fundamental for technology-agnostic lakehouse architectures. Supported by all major providers.
React Native — New architecture addressed bridge bottlenecks. It's the main recommendation for cross-platform mobile.
Svelte — No longer a niche option. It offers small bundles, strong performance, and a simpler component model. A credible alternative to React/Vue.
Trial — Worth Pursuing
Agent Skills — Modular context loading. One reason teams rethink MCP-by-default.
Mutation testing — The most honest sign of test quality. AI generates "perpetually green" tests, and mutation testing can catch these.
Sandboxed execution for coding agents — A sensible default. Use Dev Containers for ephemeral setups, and Sprites for persistent state.
Graphiti — A temporal knowledge graph for LLM memory. It tracks how facts change over time, showing 18.5% accuracy improvements in benchmarks.
LangGraph — Moved OUT of Adopt. The stateful-graph approach isn't always the right fit. Simpler patterns often yield leaner systems.
Caution — Think Twice
Agent instruction bloat — AGENTS.md files can accumulate and conflict. Models may overlook buried content. Be selective.
Codebase cognitive debt — The gap between what your system does and what your team understands is widening. AI speeds this up.
Coding throughput as productivity — Counting lines of code and pull requests creates floods of poorly aligned code. It's better to focus on first-pass acceptance rates.
MCP by default — A good CLI often suffices. Use MCP only when you need protocol-level interoperability.
Coding agent swarms — Using many agents dynamically. Successful examples relied on detailed specs and thorough tests, which is not typical in product development.
Monday Morning Takeaways
Engineer your context — use progressive disclosure, not everything at once.
Harness your agents — apply feedforward and feedback controls.
Measure what matters — focus on DORA metrics and first-pass acceptance, not just lines of code.
Security is essential — maintain zero trust, use sandboxed execution, and conduct toxic flow analysis.
Don't default to MCP — a good CLI often suffices.
The tools changed again. The principles remain the same.