Researched and written with 248,897 tokens (in 38 · out 19,247 · cache-read 78,782 · cache-create 150,830) across 5 Claude calls

Coding agents

• v2.1.290: Claude Code adds serverToolUses to the mod turn.step hook result (API-run tools like the advisor) and agentId on plugin tool.check, so hooks can tell subagent permission checks from main-session ones.
• mattpocock/skills v1.3: Adds /pr, /implement-spec (spec to subagent-built tickets) and /retro (reviews recent agent sessions, suggests repo fixes), and renames CONTEXT.md to GLOSSARY.md.
• T3 Code in-app visualization: Agents can now build dynamic UI inside the thread, with CSS style tokens matched to your theme (follow-up).
• T3 Code nightly orchestrator: User claims its orchestrator beats lab-provided tooling (subagent monitoring, roundtables, 400k+ users); anecdotal, but a signal on orchestration UX.
• trq212's HTML plan skill: A Claude Code skill that produces HTML plans with code snippets, open questions and mockups, with linting to cut common failure cases; in feedback phase.
• Anthropic made its apps 3x faster with Claude: Theo breaks down the strategies Anthropic's own agents used for the speedup, a useful template for agent-driven perf work.
• simonw/claude-system-prompts: Turns Anthropic's published claude.ai/mobile system prompts into files with git history, one commit per dated revision, so git diff shows prompt changes.
• llm-anthropic 0.30: Adds Claude Sonnet 5.5 and llm anthropic refresh, which pulls the model list from Anthropic's API so new models need no release.
• Quoting Felix Rieseberg: The "old" Cowork ran inference in the cloud and tool calls in an Anthropic-shipped local VM, which was added for capability, safety and security reasons.

Cost, tokens and routing

• OpenAI researchers' coding-agent spend: Epoch AI fits OpenAI's published weekly data (Jan to mid-Aug 2026) and finds spending doubling about once a month.
• Claude Max weekly limits fix: Reddit workflow using Opus and Sonnet sub-agents plus a review loop to stay under Max weekly caps.
• 20+ ways to make a cheap coding model act expensive: Reddit write-up of what did and didn't close the gap between cheap and expensive coding models.
• Live token optimizer for coding sessions: Community tool that trims token use during live AI coding sessions.
• codex-subscription-router: Patched macOS ChatGPT app that balances new chats across several subscriptions while pinning each thread to one account, so follow-up turns keep account-level caching.
• FastRouter.ai: Routes requests to the right LLM by cost, latency and quality.

Review, evals and safety

• ReviewBench: GitHub's open benchmark for AI code review, using real PRs, multi-source ground truth and production-aligned metrics.
• Agents write PRs faster than anyone can review: Reddit thread on review setups for agent-generated PR volume.
• 38% of agent container escapes needed no kernel 0-day: Analysis of 109 real incidents with an open dataset and defense harness; misconfiguration, not exploits, is the main route out of agent sandboxes.
• Opus 5.5 agents find room-temperature magnetic semiconductor candidates: Agent-driven materials discovery, though HN commenters stress that DFT candidates are not breakthroughs without experimental verification.

CLI/TUI and agent control rooms

• tuios: Go terminal multiplexer with BSP tiling, persistent daemon sessions across machines, and a single inbox where coding agents report their state.
• devpit: Native control room for Claude Code agents.
• Live map of files a Claude Code edit touches: Reddit tool visualizing the blast radius of each Claude Code edit.
• crosswalk: A shared space for people and their agents, starting with inbox.

Vibe-coding

• Local events site from 101 calendars: Vibe-coded PHP + SQLite site on GoDaddy shared hosting, no framework, refreshing every 6 hours.