Agentic insights for modern tech teams
CodeScene has published a case study in which coding agents refactored 300,000 lines of C over three weeks for roughly $4,000 in tokens, verified by a frame-by-frame replay harness. The agents built a playbook of codebase-specific recipes along the way. Practitioners have questioned the scope, the metric, and how much of the result depends on the harness. By Steef-Jan Wiggers
I built a synthetic multi-agent loan crew on Amazon Bedrock and made Traccia hard-block a runaway agent, redact applicant PII across every sub-agent, and export EU AI Act audit evidence. Two of my three governance policies blocked nothing - and not because I misconfigured them. Why is the useful part.
The New Personal Agent Era: 14 Architectures Redefining Personal AI For the past few...
Serdeis an amazing serialization library for Rust and it has been a huge reason why I felt productive with it for years. However already while at Sentry I got quite frustrated with some of the limitations with it but actually replacing Serde is tricky because of the might that it has in the ecosystem. Also because it’s quite hard to actually do better without also making some potentially painful compromises. Here are three examples of Serde corner cases that show poor interactions of Serde features or unexpected limitations:
I've been using OnePlus phones since the OnePlus 3T. I'm currently on a OnePlus 15 (CPH2747) running OxygenOS 16, which I've had for a few months now and am pretty happy with. I figured I'd poke at the firmware a bit to see what's in there. Specifically I wanted to know whether the clean-install firmware had anything that'd let a random app I install from outside the Play Store end up running as root. Turns out: yes! I found two separate issues which chain together into a fairly clean untrusted_app
What a refractometer taught me about observability In July, I wrote about a farmer walking a...
OpenAI launched Dots at DevDay: persistent agents powered by GPT-6 Astra that run on their own cloud computers with dedicated browsers. Unlike stateless agents, Dots work asynchronously across 4,000+ apps via the plugin ecosystem, scoping bugs, building fixes, and returning PRs with video walkthroughs without draining the user's Pro plan quota. Safety is architected via a separation of read-only proactive research from an 'auto-review' step gated by user-defined Custom Rules, preventing unauthorized actions while enabling unsupervised background work.
Pure vector search fails for agent memory because semantic similarity ignores temporal context and nuanced topic differences, such as changed opinions on architecture. A hybrid approach combining lexical and vector retrieval fused via RRF, topped with a BGE cross-encoder reranker running as an ONNX model inside Oracle, significantly improves relevance over similarity alone. The benchmark, adapted from Oracle's RAG evaluation guide, validates that in-database scoring effectively mitigates the reranker's query-time latency cost.
OpenAI's new Decisions API, built on the Luna model and announced at DevDay, directly competes with TypeSafe's Jev by returning predefined answers with confidence scores in 150ms—10x faster than GPT-6 Luna. It avoids the token cost and unreliable confidence of chat models while requiring no labeled data or retraining like custom classifiers. This makes it ideal for agent action selection, content routing, and classification tasks in production systems.
Recently launched, Amazon CloudWatch Omni is an AI-first observability platform designed to monitor, evaluate, and troubleshoot applications and autonomous AI agents in a unified environment. By Sergio De Simone