Technology Radar
January 2026 - Version 4
Tech radar refresh: January 2026
With 2025 wrapped up, it’s time to share our technology radar update based on what we’ve been building over the past year. As hot new products are on the rise more than ever, this refresh focuses on the tools and techniques that are holding up in actual day-to-day development.
Below is a description of updates from our last radar release. Check out the full radar to see everything we’re excited about.
New entries
Pyright (Adopt): Pyright has become our default Python type checker. Though we find it to be a bit slow, it works well with modern editors and legacy codebases. Looking into the future, we predict it may eventually be replaced by ty or pyrefly (currently in our “Trial” section).
FastAPI (Adopt): This is our framework of choice for building Python APIs. Though we used to use Flask frequently, FastAPI fills in all the gaps that we were finding there.
GPT Realtime (Adopt): We’ve found this to be a solid and testable voice-to-voice AI solution. It’s revolutionizing how we build IVRs in production.
TDD for AI Applications (Adopt): Whenever we are in unfamiliar territory, it’s tempting to stray from TDD when actually it can be our best ally. This concept transfers to building AI applications too. TDD is a useful tool to force clarity around behavior and expectations in the same way it does in traditional development. In fact, test-imposed guardrails are more helpful than ever.
Coding agents (Adopt): AI coding agents are clearly one of the most interesting tools changing the modern development process. Of course we should be cautious of using them with abandon, but used thoughtfully we’ve found them to be a real productivity win. Though many modern coding agents are quite useful, we are particularly fond of the AI enhancements found in our ‘Trial’ editor “Zed”, and tend to also favor Claude Code and Codex.
Freeplay (Adopt): As AI systems increase in complexity, evaluation and observability are no longer optional. Freeplay is our favorite front runner in evaluation, and testing.
AI Evaluations (Adopt): Shipping AI without evaluation is just guessing. Adding AI to our applications introduces non-deterministic behavior that would be downright risky to include without this essential metric.
RAG Search Evaluations (Adopt): Context quality is the key ingredient to RAG. Quality context ensures that the correct information is being processed at the root of the system before we layer on prompting or generation. It hardens our concept of context engineering and moves us from guessing to data driven outcomes.
MCP (Trial): Another rare standardization in the wild west of AI: MCP has given us a standardized way to communicate between AI systems without hard-coding custom integrations.
Tool calling (Adopt): Tool calling provides us with a structured way to interact with collaborators instead of relying on brittle and less predictable prompt logic. It’s become a foundational pattern for building reliable agentic systems.
Cloudflare Containers (Trial): Cloudflare Containers extend our “Adopted” Worker’s model to heavier workloads. We’ve been playing with their capacity to run stateful or compute intensive services closer to the edge.
Ty (Trial): ty is an interesting new addition in the Python typing ecosystem. It’s still early, but we’re watching its potential improvements over existing tools, particularly in the speed department (which is our main complaint about Pyrite). Don’t be surprised if you see this moved into Adopt here shortly.
Pyrefly (Trial): Another emerging type checker exploring different tradeoffs in Python typing. Once again, another contender for “Adopt”.
Maestro (Trial): Maestro is a refreshingly approachable take on mobile UI testing. As mobile development evolves, it allows us to run tests against a variety of technologies including Flutter and Kotlin Multiplatform.
Agent Client Protocol (Trial): In the “wild west” environment of AI, we have hope in the future of standardizations. For example, Agent Client Protocol is becoming the standard interface for how AI coding agents and IDEs can communicate, reducing specific integrations and vendor lock-in. It’s early, but it gives us hope for an interoperable future.
Spec driven development (Assess): At times, writing the spec is harder than writing the code, but it often surfaces issues earlier and leads to better outcomes especially regarding better tested and more testable code.
Running your own LLMs (Assess): Running your own LLMs comes with operational complexity and requires experienced data scientists to deploy and maintain. However, in certain highly regulated or legally constrained environments this practice is a necessary tool.
Promotions
Cloudflare Workers (Adopt): Cloudflare Workers are a great option when you need low latency distributed deployments with very little operational overhead. They’re especially useful for when you need edge workloads.
uv (Adopt): uv is our must-use tool for Python version and package managing projects.
Zed (Trial): This editor is delightfully fast. That alone makes it worth using. We’re also excited about their focus on collaborative work and potential to change the way we develop.
Pop! (Adopt): Pop remains a solid collaborative screen sharing option. Though it is still less feature complete and more buggy than tools like Tuple, the price tag ($0.00) and cross-platform abilities have us reaching for this as our default.
Tuple (Adopt): Ignoring the price tag, Tuple is clearly the superior screen sharing tool. It has exactly the features needed for pairing and a seamless integration into your working environment.
Demotions
Mypy (Hold): Mypy helped establish typed Python, but today we usually favor Pyright when establishing new projects. Mypy still works, but we’re no longer investing much here.
That’s a wrap
Our hope is that this snapshot helps pique your interest, spark conversations, and support intentional decisions within your projects as AI becomes a core part of our development systems. We’re always game to chat about what you’ve been exploring too!