Researched and written with 320,767 tokens (in 10 · out 38,562 · cache-read 66,228 · cache-create 215,967) across 5 Claude calls

Coding agents: cost, limits, throughput

• RT @ClaudeDevs: Background subagents in Claude Code now use less of your usage limit: status-line updates every 30s used to re-read the whole conversation; Haiku 5.5 now writes them from short notes, cutting per-update cost about 99%.
• OpenAI doubled rate limits. Is it enough?: Theo compares the value of OpenAI's new limits against before and against Claude Code subscriptions, useful for deciding where to spend your sub.
• We've consistently seen our users more than 3x their token throughput after moving to T3 Code: a vendor claim that the harness, not the model, limits how much of your subscription you actually use.
• Building a unique dashboard for a client. Cache usage is astronomical.: anecdote about heavy prompt-cache reads in a Claude Code workflow, with no body text, so it's of limited use beyond the title.

Coding agents: security and incidents

• Some OpenCode users found out their AI keys were stolen when the bill showed up: an attacker got into a legacy system OpenCode was shutting down on Oct 16 and took the OpenAI, Anthropic and Google keys users had saved in Console. Don't store provider keys on a third party's servers.
• When two coding agents talk over an open channel, they get stuck in a 40-turn polite agreement loop: a failure mode for multi-agent setups. Without a turn cap or an arbiter, agent-to-agent chat burns tokens agreeing with itself.

Coding agents: tools, releases, industry

• GitHub Copilot is now available in Xcode's coding assistant: public preview of Copilot inside Xcode's built-in assistant, which puts it in the same place as Apple's own agent integrations.
• Conductor is joining Meta: the parallel-agent worktree UI is going into Meta Superintelligence Labs to build "the next interface", so expect the product to change.
• rust-v0.167.0-alpha.3.2: Codex Rust CLI alpha tag. The captured page was only GitHub navigation, so there are no release notes to read.
• llm 0.37.1: bugfix release of Simon Willison's llm CLI, with several outside contributors.
• Thinking about shipping a version of my skills to npm/other registries: Matt Pocock's idea is registry-distributed skills, which would allow per-agent dynamic skills and version pinning in software factories.
• Ask HN: What's your current AI agentic setup?: 262 points. The common setup is one strong model in Claude Code at medium effort with few or no custom skills, plus cheap open-weight models (DeepSeek, OpenRouter) for bulk work.

Models and benchmarks

• Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding (announcement): a 65B dense model pretrained from scratch that claims US open-weight SOTA on agentic coding. HN skeptics say the claim holds only after stacking filters and omits Qwen, which scores 84.3 on Terminal Bench 2.1 against Ripple's 80.5.
• semianalysis GDPx: the best AI agent finishes 12% of real expert work: Gemini 4 Carbon lands 1.1 points behind Opus 5.5 at about a third of the cost per task, which makes score per dollar the metric to watch.

Skills, setups, and agent tooling

• garrytan/gstack: Garry Tan's Claude Code setup of 23 Markdown slash commands that play CEO, designer, eng manager, QA and security roles. MIT licensed, and worth skimming for ideas rather than adopting wholesale.
• github/spec-kit: GitHub's toolkit of templates and processes for spec-driven development with coding agents.
• The-PR-Agent/pr-agent: the community-maintained legacy open-source PR reviewer, separate from Qodo's commercial offering.
• I gave claudecode eyes on my running app: the agent clicks a button and gets back every frontend and backend line that ran, a runtime-trace feedback loop for verification. The post has no body text, so details are unknown.
• Gave 3 Claude Code UI design skills the same brief: a side-by-side comparison of UI design skills. The post has no body text, so the results are unknown.
• QualityLayer: Product Hunt launch of a tool for planning, building and reviewing software with your coding agents.

Agentic SaaS and agent infrastructure

• Render Sandboxes: "A computer for every agent", Render's entry in the agent sandbox/runtime space.
• Runway Dev on Claude Code: a plugin that lets Claude Code write Runway API integrations for video, image and audio generation into your app.
• scenario-labs/skills: agent skills that drive image, video, audio and 3D generation through the Scenario MCP. They price before spending, and expert teams drive Blender, Maya, Unreal and Unity.
• calesthio/OpenMontage: an open-source agentic video production system with 12 pipelines, 100+ tools and 700+ skill files that turns your coding assistant into a video studio.