JUXT AI Radar
Introduction
Frontier models and AI coding assistants have continued to improve at a relentless pace since our last update, and many software engineering teams are now working almost entirely within an agentic coding harness. In our work on agentic AI platforms for large enterprises, we see first-hand the same transition gradually being rolled out to other knowledge workers. The friction has shifted to the engineering and security work needed to turn their capabilities into reliable human-AI “centaur” systems. Most of this quarter’s updates are in this area.
In the Adopt ring, MCP and RAG picked up substantial new security guidance covering authentication, least-privilege tool grants, audit logging and defences against indirect prompt injection. CaMeL enters our radar in Assess as a research-level answer to the same problem. We’ve seen public failure modes this quarter: for example when Claude Code’s own source code was accidentally exposed, or headlines relating to agents with broad permissions destroying production data. We expect to be hearing many more stories along these lines this year.
The fastest-growing category is also the one we’re most cautious about. OpenClaw, NVIDIA’s NemoClaw and Moonshot’s KimiClaw are persistent agent runtimes built around the same idea: always-on autonomous agents that can control computers. The security models do not hold up to the breadth of access these systems require, and we put all three in Hold. Where they’re already in use, sandboxing and human oversight for sensitive actions are the only reasonable mitigations.
On the other hand, practices for agentic engineering are beginning to catch up. Spec-driven development enters Adopt: in our experience, structured specs shape and constrain AI implementations more reliably than ad-hoc prompts. Formal specification languages provide rigour; informal ones add little. AI-assisted code migration, LLM API gateways, red teaming and governance platforms enter Trial.
Considerations around sovereignty are coming into focus. The argument for running open weight models on your own infrastructure has shifted: EU AI Act compliance requirements and the shift from flat-rate to usage-based AI billing both reward moving inference inside the organisation. For workloads handling regulated or sensitive data, hosted APIs are no longer the obvious default.
We’re still early in this transformation, and follow it with professional and personal interest.
Henry Garner (CTO, JUXT), May 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.
Languages & Frameworks
Programming languages and frameworks that power AI development.
Tools
Software tools and utilities that enhance AI development workflows.
Platforms
Infrastructure and platform services that support AI applications.
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.
