Coding agents & CLI

• Claude Code v2.1.219 ships Opus 5 as default Opus model (1M context, fast mode $10/$50 per Mtok) plus sandbox.network.strictAllowlist for silent host-denial in sandboxed commands.
• Claude Code v2.1.218 moves /code-review to a background subagent so reviews stop filling your main conversation.
• Auto mode becomes default for Pro/Max/Team plans from Aug 14, longer autonomous runs with better dangerous-command detection than manual review.
• oh-my-pi is a Rust-core (~80k lines) terminal coding agent fork of Pi: hash-anchored edits, LSP/DAP, 60+ providers, 31 built-in tools, subagents.
• Simon Willison quotes the Opus 5 system prompt, including the Fable/Mythos export-control suspension and restoration timeline baked into model context.

Model releases & analysis

• Opus 5 lands at "Fable-level performance, half the price" per Latent Space, benchmarks edge Fable though Anthropic's own messaging is more conservative.
• Interconnects on GLM-5.2 calls it a capability step-change for open agentic models, worth watching as the open-weight alternative to closed coding agents.
• Reddit thread claims Opus 5 shows unexpected strength at photogrammetry tasks, anecdotal but signals broader spatial-reasoning gains.
• Interconnects, May roundup covers Gemini Flash 3.5, Mythos, and the open/closed balance shifting as economics of AI get real.
• Claude Fable 5 and new AI safety fables digs into Anthropic's export-control-driven safety classifier changes around the Fable/Mythos release.
• AirLLM runs 70B models on a 4GB GPU without quantization, even Kimi K3 (2.8T MoE) under 4GB, relevant for local inference cost tricks.

Agent skills & code intelligence

• addyosmani/agent-skills packages production-grade DEFINE→PLAN→BUILD→VERIFY→REVIEW→SHIP workflows as reusable Claude Code skills.
• code-graph-rag builds a Tree-sitter + Memgraph knowledge graph over a multi-language monorepo for natural-language query/edit.
• alirezarezvani/claude-skills is a 362-skill library spanning engineering to compliance, works across 13 coding agents.
• msitarzewski/agency-agents installs a themed roster of specialist subagents (frontend, community, QA) into Claude Code, Cursor, Codex, Gemini.
• harveyai/harvey-labs is an open benchmark (LAB) for evaluating LLM agents on real legal work with rubric-based scoring.

Agentic SaaS & orchestration

• stablyai/orca is a desktop/mobile ADE running Codex, Claude Code, OpenCode, or Pi side-by-side in isolated worktrees with usage tracking.
• Cloudflare OS is Cloudflare's internal Workers-based agent workspace for docs/apps/agents, now open-sourced.
• OmniRoute aggregates 43 provider free tiers (516 models) into one gateway with quota-aware fallback and prompt compression claiming 15-95% token savings.
• TencentDB Agent Memory is a team-level shared memory hub (Chat Memory, Skill, Wiki, Code-Graph) for cross-agent/cross-framework reuse.
• Toolport is a one-MCP-setup gateway aggregating tools for multiple agents.
• SnapState offers persistent state for AI agent workflows.
• Omniwork bills itself a "Creative Agent OS" for desktop creative AI agents.
• archify is an agent skill generating self-contained, animated HTML architecture/sequence/data-flow diagrams from a codebase description.

Vibe-coding & repos

• OpenChamber is an ADE that keeps agents working toward a set session goal across turns, runs one task across up to five models (multi-run/fusion), and groups large diffs into a walkthrough of ordered steps.