AI in Technology Radar Vol. 32: a review with 2026 hindsight

SECTION 1

48 AI blips: what Technology Radar Vol. 32 revealed

Volume 32 of the Thoughtworks Technology Radar, published in April 2025, featured 48 AI-related blips, just under half the total. While preparing it, its authors themselves wondered: was their attention being pulled by the surrounding hype?

With the hindsight of 2026, the answer is clear: this volume did not overestimate AI, it captured a tipping point. AI now permeates practically all of software, from how we build to what we build, and the subsequent volumes of the Radar, through Vol. 34 in April 2026, have only amplified the movement. As the field becomes embedded in everyday practice, we indeed talk less and less about "AI" as a catch-all term and more and more about specific approaches, agents, RAG, evals, exactly as the industry stopped talking about "the cloud" in general.

Rereading Vol. 32 today is a useful exercise: its groupings, AI-assisted development, models, observability, LLM frameworks, information retrieval, synthetic data, remain the best reading grid for AI as it is actually practiced across the industry, beyond the headlines. That is the grid we walk through here, flagging what 2026 has confirmed, nuanced or superseded.

SECTION 2

AI-assisted development: from assistants to coding agents

Vol. 32 broadened the landscape beyond GitHub Copilot by placing Cursor, Cline and Windsurf in Trial, all equipped with agentic capabilities. Cautious about the nascent vibe coding, the authors were enthusiastic about agentic workflows, an intuition since confirmed: in 2026, terminal-based coding agents like Claude Code and the agent modes of IDEs have become the standard tooling of development teams.

The vocabulary itself has followed: we no longer speak of assistants but of agents able to plan a task, modify multiple files, run the tests and iterate to the result. Vol. 32 captured precisely that tipping point where autocompletion gave way to supervised delegation.

### The anti-patterns flagged as early as 2025

The Radar placed complacency toward AI-generated code and replacing pair programming with AI in Hold, while highlighting AI-friendly code design (Assess) and the idea that AI should deliver team gains, not just individual ones. It was also wary of local coding assistants, limited by their context window and their inability to trigger tool calls, a judgment that hardware progress has only partially overturned since.

### The 2026 hindsight: putting coding agents on a leash

These warnings have grown in scale: Vol. 34 devotes an entire theme to putting coding agents on a leash and warns about cognitive debt, the widening gap between humans and systems in which a massive share of the code is AI-generated. The underlying message has not changed, it has hardened: vigilance over generated code, yesterday a simple good practice, has become a full engineering discipline, with systematic reviews, tooled guardrails and a deliberate return to fundamentals.

@cite:a-quelle-vitesse-les-assistants-de-codage

SECTION 3

AI models: DeepSeek R1 and the end of 'bigger is better'

On the model side, the event of Vol. 32 was the emergence of DeepSeek R1 (Assess). Beyond the financial and geopolitical noise, the DeepSeek team's work was highly innovative: its efficiency gains signaled that there are alternatives to the somewhat blunt "bigger is better" philosophy. The Radar also praised Claude Sonnet (Trial) for its coding results, reasoning models (Assess), small language models that run on edge devices (Trial) and ModernBERT, the encoder successor to BERT for natural language processing tasks.

### Fine-tuning, distillation and the race for efficiency

DeepSeek had sharpened interest in model optimization: the volume flagged distillation (Trial), which transfers knowledge from a large model to a smaller, more efficient one, and torchtune, a Python library for post-training LLMs.

The hindsight of 2026 validates the analysis. Reasoning has become a standard capability of frontier models, current generations such as Claude Sonnet 5 or the GPT-5.6 family reason natively, no longer as a separate category, while the race for efficiency, driven by small specialized models, distillation and quantization, has established itself as the sector's dominant economic axis. Exactly what the DeepSeek breakthrough heralded.

@cite:demystifier-deepseek

SECTION 4

Observability, evals and guardrails: AI put to the test of production

As AI systems moved into production, observability and the evaluation of their reliability became critical. Vol. 32 promoted NeMo Guardrails from Assess to Trial after broad adoption on Thoughtworks teams' projects, and flagged alternatives to Langfuse: Arize Phoenix, Weights & Biases Weave and Helicone, all in Assess, along with Humanloop, deemed particularly useful in regulated domains where risk management is critical.

This proliferation reflected a realization: an LLM system cannot be signed off like deterministic software. Without detailed traces of prompts, responses, latencies and costs, there is no way to diagnose a quality drift or a bill explosion, the tools cited filled precisely that gap.

### From observability to interpretability

The volume already distinguished interpretability, validating outputs to guarantee their reliability and accuracy, from observability, pointing to Gemma Scope as a tool worth examining.

This is arguably the most prescient section of Vol. 32: in 2026, evals have become a central engineering discipline, Vol. 34 makes it the heart of its theme on evaluating technology in an agentic world, and no serious agentic system goes to production without a versioned evaluation suite, LLM tracing and tested guardrails, just as classic code ships with a test suite. The teams that understood this as early as 2025 approached the agentic turn with a measurable head start.

@cite:ai-evals-pour-mcp-dans-aiops-comment-evaluer-et-ameliorer

SECTION 5

LLM frameworks and RAG: from PydanticAI to GraphRAG

Without reaching the scale of the mid-2010s JavaScript framework boom, the LLM framework ecosystem was growing fast. After long-standing reservations about LangChain, Vol. 32 highlighted PydanticAI, a lightweight Python agent framework with built-in structured output handling and a graph-based abstraction for complex agentic workflows. 2026 has settled the matter: agentic workflows are now built on frameworks that have reached stable versions, LangGraph 1.0, Microsoft Agent Framework, the unified successor to AutoGen and Semantic Kernel, and interconnect through the Model Context Protocol, now the de facto standard for connecting agents to tools.

### The R in RAG: knowledge graphs to the rescue

Information retrieval was a theme of the volume, the R in RAG: GraphRAG (Trial) and Fast GraphRAG leveraged knowledge graphs to improve retrieval accuracy, alongside Graphiti, which builds dynamic, time-aware knowledge graphs, and VectorChord, a high-performance PostgreSQL extension for vector search. These techniques have since entered the everyday toolbox of data teams: hybrid retrieval combining graphs, vectors and keywords has become the reference architecture of retrieval systems, including for feeding agent memory.

The contrast with 2025 fits in one word: consolidation. Where Vol. 32 listed candidates, 2026 lines up standards, interconnection protocols, stable frameworks, documented retrieval patterns, on which teams can build without fearing obsolescence within three months.

SECTION 6

Synthetic data and AI readiness

Finally, the volume featured Tonic and Synthesized (Assess), in service of what Thoughtworks calls AI readiness, having usable, representative and compliant data before building anything on top of it.

The market has followed: between regulatory requirements on personal data and models' appetite for quality training and test data, synthetic generation has settled into test pipelines, the training of specialized models and the validation of agentic systems, where testing an agent on real production data is rarely acceptable.

For QA and data teams, the most concrete use remains generating test datasets faithful to production distributions but free of personal data, a prerequisite that has become unavoidable with the tightening of GDPR controls and the European AI Act's requirements on training data.

SECTION 7

From Vol. 32 to Vol. 34: reading the Radar with 2026 hindsight

Saying the AI landscape evolves at a dizzying pace is trite; measuring it is less so. Two volumes later, Vol. 34 of April 2026 structures its themes around evaluating technology in an agentic world, securing permission-hungry agents, keeping coding agents in check and returning to engineering fundamentals in the face of cognitive debt, the logical, hardened continuation of Vol. 32's signals.

The methodological lesson is worth keeping: the Radar matters less for each individual blip than for its trajectories. Tracking a tool's movement from Assess to Trial, or the appearance of a Hold on a fashionable practice, says more about a technology's real maturity than any marketing announcement. AI remains a complex, fragmented field that has not finished settling down, if it ever does.

This hindsight also recalls an obvious truth: no volume predicts the future. Vol. 32 anticipated neither the speed of MCP's standardization nor the scale of the agentic turn of 2026; it had, however, identified all of its preconditions, model maturity, evaluation tooling, cost pressure and the engineering discipline needed to take advantage of it.

At Adservio, we use this trajectory-based reading to advise our clients on their AI tooling choices: adopt what has proven itself in production, experiment with what is rising within an evaluation framework, and avoid the anti-patterns the Radar documents volume after volume. This tooled watch avoids two symmetrical pitfalls: equipping too early on immature technologies, or waiting so long that the competitive advantage has evaporated.

Disclaimer: The statements and opinions expressed in this article are those of the author(s) and do not necessarily reflect the positions of Adservio.

FAQ

Frequently asked questions