Coding agents & LLM mechanics

• Claude Fable 5.1 and Claude Mythos 5.1: Anthropic's new flagship coding/research model ships as GA (Fable) and limited-safeguard (Mythos) variants; $0.25/Mtok cache reads reportedly cut agentic-workflow costs 25-45%, though community reports Fable's stricter alignment checks make it less reliable in practice.
• v2.1.257: Claude Code makes Fable 5.1 the default model (1M context, $10/$50 per Mtok) and adds configurable time format/timezone for turn-end clock and transcript timestamps.
• Claude Fable 5.1 made me a really nice animated pelican: Simon Willison runs his standard SVG-pelican benchmark against the new model for a quick qualitative read.
• Claude Fable 5.1 is generally available in GitHub Copilot: Fable 5.1 lands in Copilot, positioned for long-horizon autonomous coding and knowledge-work tasks.
• Kilo Code for JetBrains: fully native, open-source coding agent for JetBrains IDEs, widens agent-in-IDE options beyond VS Code/Copilot.
• The efficient frontier of LLM inference: Baseten's framing of cost/capability tradeoffs in inference engineering; community pushback notes it omits quality/intelligence as a third axis beyond throughput-latency.
• Building Autonomous Goal Loops That Deliver: argues test-pass-only agent loops miss failures outside known test coverage; light on concrete scoring methodology per community critique.
• Show HN: Running 104GB Qwen3.8-Flash-Next on 48GB Mac at ~12 tok/s: single Swift binary streams a 125B-param 4-bit MoE model from SSD to run on memory-constrained Macs, speaks Ollama/OpenAI APIs so existing tooling works unchanged.
• Copilot code review can now approve pull requests: Copilot can be authorized (admin-gated, off by default) to sign off on PR approval, a step toward autonomous merge pipelines.
• addyosmani/agent-skills: packaged production-grade engineering skills (define/plan/build/verify/review/ship) meant to keep coding agents consistent across development phases.
• I built a 4.7k-star open source tool with Claude Code without knowing Python: workflow writeup on avoiding "AI slop" output while vibe-coding an entire OSS project.
• We hit skills sprawl at 33 skills in a 6-person team: practical governance lesson on consolidating a team's Claude skill library before it becomes unmanageable.
• Fable 5.1 vs Claude Code + model routing: direct comparison of using Fable 5.1 versus letting Claude Code's model router pick.
• I gave Fable 5.1 one prompt for a walkable 3D Berlin in raw WebGL2: single-prompt, four-hour build of a no-library WebGL2 world, capability demo for long-horizon generation.
• Grok Bot but for Claude Code: third-party bot integration bringing Grok-style automation into Claude Code workflows.

Vibe-coding & repos

• PRs NOT Welcome: top AI-native OSS projects (Flue, tldraw) now reject external PRs, mostly AI-generated, running their own agents to write changes instead, a break from GitHub's 18-year open-PR default.
• Gitlawb/openclaude: open-source coding-agent CLI supporting OpenAI-compatible, Gemini, Codex OAuth, and Ollama backends under one terminal-first workflow, an alternative to vendor-locked CLIs.
• VoltAgent/awesome-design-md: collection of DESIGN.md files from popular brand design systems to drop into a project so agents generate UI matching that brand's language, extends Google Stitch's DESIGN.md concept.
• browser-use/video-use: open-source video editing via chatting with Claude Code, drop raw footage in and get a color-graded, filler-trimmed final.mp4 out.

Agentic SaaS

• How canvases make agentic workflows visible, steerable, and cost-efficient: argues chat loses agent work in the scroll, canvases give visibility and control over long agentic runs.
• Keiki: build one customer-facing AI agent and deploy it across every channel.
• Cosmic Agent Plugins: connects Cosmic agents to any service via an MCP server.
• Naseem: native macOS AI agent that performs real work directly on the machine.
• ThunderPhone: platform for building reliable AI phone agents from 2c/min.
• Murmell: shared async workspace built for AI agents to keep working while you close your laptop.

next: write vault/Daily/2026-09-02.md

• no frontmatter stats (tokens/cost/signal IDs) available here, pipeline usually computes those [fact]
• 25/25 signals covered across 3 theme groups [fact]