Explore Technology Radar Vol.33 through the lens of AI, security, infrastructure and more

Hello!

At the start of November we launched Thoughtworks Technology Radar Vol.33. We’ve had a great response to it so far, and have hugely enjoyed hearing your perspectives on what featured and where we placed it on the Radar. 

However, we know there’s a lot to go through. With 114 blips covering everything from AI development to post-quantum cryptography, getting that meso-level view of what’s happening and what’s important in different fields can be tricky.

So, in this email we’ll take a look at a selection of highlights from the recent edition of the Technology Radar by fields — it’s not exhaustive, but you can use it to go back and explore vol.33 further.

Best Wishes,

Richard Gall

Technology writer and content strategist, Thoughtworks

AI

Nearly two-thirds of the blips on this edition are AI-related so it’s hard to capture everything. However, a couple of things are worth noting — first, the Model Context Protocol (Platforms/Trial) is now on the Radar and, more importantly, it runs throughout the whole report, highlighting the extent of its impact on the space. Second, the agentic boom appears to be in full swing. Indeed, MCP is likely having an impact here, but we can also see it in the emergence of other protocols like Agent2Agent (Platforms/Assess) and AG-UI (Platforms/Assess) as well as agent development frameworks such as LangGraph (Languages and Frameworks/Adopt).

Other AI-related blips worth noting:

• This edition featured a number of technologies aimed at AI observability and testing, including Datadog LLM Observability (Platforms/Trial) and LangSmith (Platforms/Trial)
• We moved LLM as a judge (Techniques/Assess) back from Trial to Assess due to an increasing awareness of complexities and risks. “When LLM as a judge is used in training pipelines for reward modeling,” we write in the Radar “it can introduce self-enhancement bias — where a model family favors its own outputs — and preference leakage, blurring the boundary between training and testing.”

AI-assisted Software Engineering

Because AI features so extensively we also wanted to spotlight one specific element — AI assistance in software development. This was a particularly rich area for us, and reflects the work we’ve been doing to establish what works and what doesn’t. We’re particularly pleased to be able to put using generative AI to understand legacy codebases (Techniques/Adopt) into Adopt, but we’re also encouraged by signs of success in forward engineering with generative AI (Techniques/Assess) and AI for code migrations (Techniques/Trial). 

We’re seeing agentic coding evolving rapidly and particularly like Claude Code’s (Tools/Trial) capabilities — however, it’s the techniques and practices that are driving things forward as much as the tooling. Context engineering (Techniques/Assess) is becoming a significant term in the industry, but it’s also worth noting the more specific practices and techniques like teams of coding agents (Techniques/Assess), anchoring coding agents to a reference application (Techniques/Assess) and curated shared instructions for software teams (Techniques/Adopt) that highlight that beneath the high-level help there’s some thoughtful experimentation happening.

More AI-assisted software engineering blips worthy of a mention:

• We can’t mention context engineering without also highlighting spec-driven development (Techniques/Assess), an emerging AI-assisted development approach that is oriented around the specification at the heart of a project.
• N8n (Platforms/Trial) also featured, although it’s somewhat unique — it combines AI-assistance with low code capabilities.

Data

Many facets of the data field are today foundational to the demands of AI. This is certainly reflected in this edition of the Radar. Amazon S3 Vectors (Platforms/Assess), for example, promises “AI-ready storage for development at any scale” while Cleanlab (Tools/Trial) is a helpful (and open-source) tool that helps tackle issues and improve training data. We’re also intrigued by seeing some success with using GraphQL as a data access pattern for LLMs (Techniques/Assess).

More data blips worth your attention:

• We highlighted some problems with standalone engineering teams (Techniques/Hold) due to the potential for silos. Integrating data engineering into cross-functional teams is, we think, a superior approach.
• Delta Sharing (Platforms/Trial), a protocol designed to make sharing data across different platforms more secure and efficient, also stands out as valuable, particularly for teams and organizations grappling with fragmented systems. 
• Graft (Platforms/Assess) a transactional storage engine that enables strongly consistent and efficient data synchronization across edge and distributed environments.

Security

The rise of MCP, which, despite the hype around it, has some security challenges, makes tools like MCP-Scan (Tools/Assess) welcome, while techniques such as toxic flow analysis (Techniques/Assess) are valuable as agentic systems open up new opportunities for data exfiltration. Google’s Secure AI Framework (SAIF) (Techniques/Assess) also features on this volume of the Radar and promises to be a helpful guide for navigating the security challenges of AI systems.

More security blips you may have missed:

• We moved continuous compliance (Techniques/Adopt) forward to Adopt highlighting that we now see it as something that should be fundamental to how organizations think about security.
• We also moved Presidio (Languages and Frameworks/Trial), an SDK for anonymizing sensitive data from Assess to Trial.
• A post-quantum world may feel some way off, but we like that Java is thinking about some of the potentially significant risks of the future with Java post-quantum cryptography (Languages and Frameworks/Assess).

Infrastructure and Cloud

Kubernetes came up a lot in the discussions for this edition — it didn’t quite make it into a theme but you can see it crop up in a number of blips. For instance, we feature Karmada (Platforms/Assess), a platform for orchestrating workloads across multiple Kubernetes clusters, clouds and data centers, Kueue (Tools/Assess), a Kubernetes-native controller for job queuing that manages quotas and resource consumption and Crossplane (Languages and Frameworks/Trial), a framework that can help extend K8s clusters.

More infrastructure and cloud blips worth exploring:

• Barman (Tools/Trial) an open-source tool for managing backups and disaster recovery of PostgreSQL servers.
• The intriguing Oxide (Platforms/Assess) which provides prebuilt racks with compute, networking and storage, running fully integrated system software.
• We also brought service mesh without sidecar (Techniques/Assess) back to the Radar — after more than three years — as a result of Istio ambient mode.

App and Web Development

This volume featured a number of interesting development frameworks including Fastify (Languages and Frameworks/Adopt), Phoenix (Languages and Frameworks/Trial) and Nuxt (Languages and Frameworks/Trial). It’s also worth mentioning Tauri (Languages and Frameworks/Trial) too — built on Rust, it allows you to develop high-performance desktop applications using a single web UI codebase.

More development blips to explore:

• We put pnpm (Tools/Adopt), a package manager for Node.js into Adopt.
  
• OpenFeature (Platforms/Assess) is a useful abstraction layer for managing feature flags worth exploring.