New York

October 15–17, 2025

Berlin

November 3–4, 2025

The importance of supervising your AI coding agents

And 3 other themes from the latest Thoughtworks' Technology Radar.
April 02, 2025

You have 1 article left to read this month before you need to register a free LeadDev.com account.

Estimated reading time: 4 minutes

Supervised coding assistants are becoming a core part of software development, according to the latest Thoughtworks Technology Radar.

Every six months, 20 or so senior technologists at the software consultancy Thoughtworks get together to compile the latest Technology Radar, a snapshot of the various tools, techniques, platforms, languages, and frameworks its consultants have encountered in the wild, each of which is represented by a blip on the radar.

During those discussions, some common themes emerged. LeadDev spoke to Ken Mugrage, a principal technologist at Thoughtworks, about how supervised agents are the future of developer teams, and the rapid evolution of observability.

Supervised agents

Generative AI-powered coding assistants are developing fast, and while fully autonomous AI engineers are likely a long way from being a reality, tools like Cursor, Cline, and Windsurf are starting to look more like supervised assistants than a glorified autocomplete function. These tools are able to generate code snippets, modify code, update tests, and proactively fix linting and compilation errors.

“We will see a future where code is co-written by humans and AI agents,” GitHub CEO Thomas Dohmke predicted in January. “The human sets the overarching goal, determines constraints, ensures ethical considerations, and divides the work into small chunks that can be handled by the state-of-the-art model, while the AI agent takes on the grunt work of writing, testing, and refining large swaths of code.”

For Mugrage at Thoughtworks, the rapid development of agentic or chat-oriented programming (CHOP) is allowing developers to overcome the cold start problem at the beginning of a project, but careful supervision of the agents is more important than ever in professional environments.

What engineering leaders need to know: Don’t get complacent with AI-assisted code and maintain discipline around basic code quality standards. “We have guidelines around test coverage and sensible defaults for a reason,” Mugrage said. “It’s important that people are trained on those things and understand what they are and agree with them, frankly.”

Evolving observability

As engineers look to integrate large language models (LLMs) into their workflows more, the ability to observe and infer the behaviour of those systems becomes more critical than ever. Tools for doing this, like Weights & Biases Weave, Arize Phoenix, Helicone, and HumanLoop, all popped up on the April edition of the Technology Radar as a result.

Organizations’ observability needs are also evolving at pace, driving Honeycomb founder Charity Majors to declare the emergence of observability 2.0. “Our systems are exploding in complexity along with power and capabilities,” Majors wrote. “The idea that developing your code and operating your code are two different practices that can be done by two different people is no longer tenable.”

What engineering leaders need to know: Majors’ put it best when she wrote: “You can’t operate your code as a black box, you have to instrument it.” Only then can engineering leaders build confidence in the outputs.

Putting the R in RAG

Assessing the quality of generative AI outputs is a major concern for engineering leaders entrusted with incorporating the technology into their products and engineering toolchains. Retrieval-augmented generation (RAG) is an important technique for engineering teams looking to fine-tune their outputs, and is already a space seeing significant innovation.

Thoughtworks plotted a whole host of trends in this domain, from corrective RAG – which dynamically adjusts responses based on feedback or heuristics –  to Fusion-RAG – which combines multiple sources and retrieval strategies for more comprehensive and robust responses. There is also Self-RAG, which avoids the retrieval step altogether, fetching data on demand, and FastGraphRAG, which aids understandability by creating human-navigable graphs.

What engineering leaders need to know: This is an immature and fast-moving space, with no standard approach, making assessment and selection tricky. Listen to your engineers and pick the best option for your needs, rather than trying a one-size-fits-all approach.

Taming the data frontier

The complexity of data ecosystems is a major concern for engineering leaders, especially those being tasked with developing internal generative AI projects.

“This is the age-old garbage in, garbage out problem,” Mugrage said. “This is fundamental to doing business in this generation. You have to have your data in order.” 

As a result, vector database tools and analytics products like Metabase both appeared on the latest Technology Radar, as well as frameworks like data product thinking.

What engineering leaders need to know: The foundations need to be solid to truly adopt generative AI, and that includes unstructured data. This means enforcing discipline around metadata management and having a good feel for when the data is good enough to build on.


Final thoughts

“The true reflection of the discussion in the room was that AI touched almost everything,” Mugrage said. 

Like it or loathe it, engineering leaders are being forced to reckon with the impact of generative AI across a variety of areas, but a lot of the advice here centers on getting back to basics: know what good looks like, be disciplined, and build confidence in what you’re adopting and building.