Software Tech Radar

Google I/O Agentic Tooling Releases

• 🔥 Android CLI for Agentic Coding: Google released Android CLI to enable AI coding agents like Claude Code...

Teams are moving AI agents into production with three recurring patterns.

• Hybrid memory stacks automatic transcripts, curated files, and vector...

• Test-driven multi-agent systems generate full-stack web applications directly from requirements, advancing from runnable prototypes to shippable...

Google is pivoting from chatbots to production-grade autonomous agents through coordinated launches.

• Gemini 3.5 Flash delivers frontier...

The PuppyChatter Framework directly addresses core difficulties in developing AI applications that rely on Large Language Models, prioritizing usable and flexible tooling for engineers.

Fresh research from AI labs this month zeroes in on practical breakthroughs for developers.

• Agentic AI advances enable more autonomous, goal-driven...

Google is advancing agentic coding by launching an Android CLI that integrates directly with AI agents like Claude Code and OpenAI’s Codex.

• Lets...

A concise summary titled The last six months in LLMs in five minutes has drawn strong interest, racking up 713 points on Hacker News and delivering a rapid overview of recent model developments for busy engineers.

Andrej Karpathy's move to Anthropic delivers elite research firepower for Claude model advancement, blending his OpenAI and Tesla experience with...

Gemini is guaranteed to top every benchmark like always, but the key uncertainty remains whether it will be a genuinely useful model or simply another benchmaxxed one optimized for scores rather than real-world performance.

GPU designs optimized for dense workloads are hitting limits with sparse frontier LLMs, where memory bandwidth—not FLOPS—dominates performance and...

PolyAI is giving developers direct access to the same conversational AI infrastructure powering major enterprises like Marriott and FedEx.

• Poly...

Developers are seeing the first wave of tools built from the ground up for autonomous agents rather than humans.

• Programming models: A new language...

Lance presents a unified approach to multimodal modeling through multi-task synergy. Discussion continues on the paper page.

CompactAttention introduces Block-Union KV Selection to accelerate chunked prefill in transformers, delivering a targeted optimization for LLM inference that software engineers working on model serving should watch.

AI Coding Model Updates

• 🔥 Cursor Composer 2.5 Release: Cursor ships Composer 2.5, an AI coding model built on Kimi K2.5 that matches Opus 4.7...

Traditional vector RAG struggles with chunking and embedding limitations as long-context LLMs reshape retrieval needs.

• Vectorless methods combine...

Cursor's new in-house model delivers strong coding performance at much lower cost.

• Performance: Matches Opus 4.7 on some benchmarks and GPT-5.5
  -...

Anthropic's acquisition of Stainless pulls a key SDK generator used by OpenAI, Google, and others behind closed doors, winding down hosted access...

AI agents shift from passive response to active execution, performing tasks with minimal oversight.

• Planning: Break complex goals into step-by-step...