Researched and written with 308,070 tokens (in 10 · out 34,004 · cache-read 65,545 · cache-create 208,511) across 5 Claude calls

Coding agent quality and benchmarks

• SlopCodeBench breakdown: a benchmark ran 15 coding agents on tasks that evolve across checkpoints, and code quality got worse in most runs, which matters for any long-lived agent-edited repo.
• ADEMark, the first real ADE benchmark: 16 agentic dev environments (T3 Code, Cursor, Capy, Paseo and others) run on the same tasks and machine with every frontier model, measuring tokens, speed and resource weight, so harness choice becomes a measurable variable.
• Superlite: hand an agent a spec: an autonomous Claude Code build workflow that skips constant approval prompts; commenters doubt Opus 5.5 needs Superpowers-style scaffolding and ask how parallel agents avoid editing each other's work.
• Michigan EECS 498-016 Applied Agentic Software Engineering: a 15-week course where students build their own tool-call, agent-loop and permission layers, then run at least 8 tasks 3 times each for success rate, cost and latency, with open-sourced slides and demos.

Tokens, context and cost

• Pre-commit hook capping files at 5,000 tokens: Matt Pocock's prompt blocks staged agent-authored files over bytes/4 tokens and puts split instructions in the error message, a cheap way to keep files agent-friendly.
• Most token-efficient way to orchestrate Claude Code models?: a thread on model routing inside Claude Code for token cost.
• Splitting planning from clicking: a desktop agent got much faster with a frontier model planning and a small model executing the actions, the same planner/worker routing pattern applied to computer use.
• Hybrid Reasoning Agents: agents reason in latent space and act through text; the open questions are whether it improves accuracy or only cuts tokens and latency, and whether hidden reasoning is harder to monitor.

Observability and data leakage

• shim-cli: shows what your coding agent sends upstream and masks what it can.
• If you don't run your own LLM proxy, how do you know what's going on?: argues for owning the proxy layer; commenters favour a LiteLLM proxy with virtual keys so provider keys stay out of sandboxes, with OpenRouter as the zero-setup option.

Claude Code and agent runtimes

• Claude Code v2.1.303: bug fixes and reliability only, no new features.
• Claude Code desktop UI session cards: the desktop UI can now propose session cards and spawn new sessions, which eases multi-session workflows.
• Agent Code 2.0 Development Preview: a major update to Agent Code, now in preview.
• Anthropic rebuilt git itself: Theo reports Anthropic was bottlenecked by GitHub and rebuilt git, and argues heavy agent users will hit the same scaling wall.

Agent control surfaces

• Claude Code sessions as a pixel town: an open-source, local-first view that shows each Claude Code session as a building.
• Claude workers in a pixel office: free, open-source workers turn GitHub issues into tested pull requests.
• Claude Code answering phone notifications: the open-source app that lets Claude Code use your phone now replies to messages when a notification arrives.
• One phone and browser app for all your agents: a single mobile and web client in place of one app per agent.

Vibe-coding and agent frameworks

• DecisionAct: an open-source Python agent framework that uses decision models instead of LLM-generated replies.
• Plain Claude designing real hardware: an open-source Pocket recorder designed down to CNC drawings in a week with no specialised tooling.