This week

Agents take the wheel

A week of agentic firsts — coding agents shipping production kernels, computer-use models going open-weight, and AI labs racing to own the full developer stack. The era of AI as a passive tool is ending fast.

Edition #54

Sunday 22 March 2026

7 stories · 3 min read

Open weights arrive Computer use for all now The agents awaken

AI & Models

Simon Willison describes using LLMs — primarily Claude Opus 4.6 — to generate detailed personality and behavioral profiles of HN users from their last 1,000 public comments, fetched via the Algolia HN API.

Products & Open Source

Floci is a new MIT-licensed local AWS emulator positioned as a drop-in replacement for LocalStack Community Edition, which sunset in March 2026 by adding auth requirements, dropping CI support, and freezing security updates. Atomic is a self-hosted personal knowledge base (Tauri desktop app + Docker server) that chunks and embeds markdown notes using sqlite-vec for semantic search, then surfaces them via force-directed canvas, LLM-generated wiki synthesis, and an agentic RAG chat interface. On March 19, 2026, attackers used retained credentials (from incomplete rotation after a February breach) to publish a malicious Trivy v0.69.4 release and overwrite 76/77 tags in trivy-action and all 7 tags in setup-trivy with credential-stealing malware.

Infrastructure & Engineering

Simon Willison's guide covers practical Git workflows for working with coding agents, including branching strategies, commit hygiene, and using Git history as a safety net when agents make mistakes. Tiny corp is now shipping the Tinybox, a line of deep learning workstations built around their tinygrad framework. Tooscut is a browser-based professional video editor leveraging WebGPU and WebAssembly for native-level performance.

Predictions

Signal-grounded, self-evaluating

highProducts

A major AI lab will acquire or partner with a developer tooling company

within 4 weeks

Signal: 3 sources covering devtools acquisitions, accelerating over 5 days

mediumModels

Local inference will match GPT-4 quality on coding benchmarks

within 3 months

Signal: Flash-MoE running 397B on laptop + 4 similar efficiency papers this week

moonshotModels

A major open-source computer-use agent will surpass proprietary alternatives on at least one benchmark

within 6 weeks

Signal: Holotron-12B release + 2 open computer-use papers in same week