JUXT AI Radar – Q4 2026

An engineer's guide to the AI landscape from JUXT's CTO & AI Chapter members

JUXT AI Radar

Introduction

This quarter, the upheaval AI has wrought on software engineering over the past couple of years reached another discipline: mathematics.

In early October, OpenAI released hundreds of AI-generated proofs, many formalised using the Lean proof assistant. The community would likely have celebrated any mathematician who cracked even just one of these problems; yet here they arrived all together, solved mostly autonomously by one of OpenAI’s unreleased models. Mathematicians did not celebrate. The mood was unease, bordering on anger.

A proof is more than its conclusion: much of its value is in the route to it, in the understanding built along the way. Eminent mathematician Terence Tao has likened the AI shortcut to “taking a helicopter to drop you off at the [summit]”, where you reach the result but “miss all the benefits of the journey itself”. Software has long allowed the same thing, code that runs even when no one quite understands how, but agents make it the default. They hand over a plausible result while the comprehension necessary to change or challenge it is skipped. Some call what’s left “epistemic debt”, or in its starker form “cognitive surrender”.

One remedy is to invest time in specification. We see formal specification languages, our own Allium among them, as the way to mitigate epistemic debt in software development. A formal specification, unlike an informal one written in markdown, offers certain guarantees that the design hangs together, much as Lean does for a proof. It also points you to the parts of the code that need strong tests to validate what an agent produces. Writing one keeps you engaged with the essential complexity of the problem, and hands the machine only the incidental complexity of the implementation. The closer a field can state formally what “correct” means, the more it can safely delegate.

This quarter the radar includes several other techniques for guiding and constraining AI output. Harness engineering, context engineering, evals as a release gate, orchestrator-worker agents and AGENTS.md and steering files all enter Trial, and Claude Skills moves to Adopt. In Languages & Frameworks, the first-party vendor agent SDKs enter Trial. Spec-driven development stays in Adopt.

Checkers like the Lean proof assistant, or automated tests, verify that an agent’s output is coherent, whereas guardrails limit what the agent is permitted to do. Some very public guardrail failures have occurred recently. A swarm of OpenAI’s evaluation agents tasked with beating a benchmark escaped a test sandbox and reached Hugging Face’s production systems seeking the benchmark’s answers. The radar adds blips informed by this and other failures: agent code sandboxes, MCP gateways, AI supply-chain security and agentic vulnerability discovery enter Tools, and agent identity enters Techniques.

Even as the frontier labs keep impressing on the world—and on potential investors—how risky their models are, regulation is being deliberately kept loose. The EU’s Digital Omnibus pushed the AI Act’s high-risk obligations to December 2027 to help keep European AI competitive. It kept the duty to mark AI-generated content. AI content provenance enters Techniques and AI governance platforms moves up to Trial.

For now, the only dependable constraints are the ones teams build for themselves.

Henry Garner (CTO, JUXT), October 2026

Radar Overview

Our radar is organised into four main categories, each containing technologies evaluated across four adoption levels:

  • Adopt: Technologies we recommend using now
  • Trial: Worth exploring for new projects
  • Assess: Keep under observation
  • Hold: Not recommended for new projects

Categories

Techniques

AI methodologies and practices that shape how we build intelligent systems.

Adopt, Trial, Assess, Hold

Languages & Frameworks

Programming languages and frameworks that power AI development.

Adopt, Trial, Assess, Hold

Tools

Software tools and utilities that enhance AI development workflows.

Adopt, Trial, Assess, Hold

Platforms

Infrastructure and platform services that support AI applications.

Adopt, Trial, Assess, Hold

Contributing

This radar represents our current viewpoint and will be updated regularly. We welcome feedback and suggestions from the community, you can reach us on LinkedIn, BlueSky and via email. Each technology entry includes detailed reasoning for its placement, helping you make informed decisions for your AI projects.

Get industry news, insights, research, updates and events directly to your inbox

Sign up for our newsletter