Microsoft's CodeAct: A New Approach That Lets Agents Call Tools in Batches

CodeAct, a new feature announced by Microsoft, fundamentally rethinks how AI agents invoke tools. Conventional agents repeat a cycle of 'emit one tool call as JSON → execute → receive the result and reconsider' — an inefficient structure in which even comparing the weather across two cities requires calling get_weather twice, with a full round-trip of model inference running each time.

CodeAct reduces these per-step round trips by letting the model describe and execute multiple tool calls at once through code. This brings agent behavior closer to natural, 'programming-like' conduct, and opens up the possibility of completing complex tasks in fewer steps. As a fundamental paradigm shift in agent design, its future impact will be worth watching.

C2PAVerify: An Open Source Tool That Distinguishes AI Images from Camera Photos

C2PAVerify, released on GitHub by developer Wavesonics, is an open source application that determines whether a photo is AI-generated or captured with an actual camera. It leverages metadata verification based on the C2PA (Content Provenance and Authenticity) standard, making it possible to technically verify the provenance of an image.

As AI-generated images proliferate, judging the authenticity of content has become a key media literacy challenge. This tool offers a practical approach to the problem, and the fact that it is open source is also meaningful from a transparency standpoint. Since platforms supporting C2PA are still limited at this point, it is not a universal solution, but it deserves attention as an early example of verification infrastructure.

Running Local LLMs on an Intel Arc B70: llama.cpp SYCL Benchmarks

Benchmark results for the llama.cpp SYCL backend on an Intel Arc B70 GPU were shared in the LocalLLaMA community on Reddit. Processing speeds (tokens/sec) were measured across multiple models, including Gemma4 12B Q8_0, Gemma4 26B.A4B Q8_0, Gemma4 E2B Q8_0, and Qwen3.5 27B Q8_0.

Notably, Gemma4 E2B (4.65B parameters) reached a generation speed of 109 t/s, confirming that lightweight models can run at practical speeds. Qwen3.5 27B, on the other hand, managed only 15.4 t/s, exposing the GPU memory and compute constraints that become apparent with larger models. The poster assessed it as 'it works, but there's room for improvement,' suggesting that local LLM environments on Intel GPUs are still a work in progress.

AI Coding Tool Automation Accelerates: Kiro and Devin

Automation in AI coding assistants is advancing yet another step.

For **Kiro**, a configuration example shared on Qiita shows how to use steering files (config files under `.kiro/steering/`) to achieve a 'fully automatic' setup: proceed without asking questions, edit files without approval, and even fix errors automatically. It is provided as a collection of copy-paste-ready config files, making it highly practical.

Meanwhile, for **Devin for Terminal**, it was demonstrated that beyond interactive use, `devin -p` (non-interactive mode) lets you embed it in scripts and run tasks in batch fashion. This is a useful approach for integrating it into CI/CD pipelines and automating routine tasks.

Both of these illustrate a shift away from the assumption of 'a human conversing in the moment' toward 'autonomous completion,' showing that AI coding tools are increasingly being absorbed into infrastructure.

The Structural Limits of AI-Driven Development: LLM 'Reasoning Deadlock' and Hallucination Chains

Around the same time, an article series analyzing the structural limits of LLMs in AI-driven development has also been drawing attention. Chapter 3 introduces the concepts of 'reasoning deadlock' and 'hallucination chains,' explaining the mechanisms by which LLMs hit dead ends.

The article points out that the quality of AI-generated code structurally depends on the limits of human cognition, and that without review processes (such as a recommended three-person review setup), quality can collapse. While automation of AI coding tools moves forward, keeping an awareness of these limits seems like a healthy balance.

---