Claude Code & Anthropic

• Claude Code v2.1.219 adds Opus 5 as default (1M context, $10/$50 per Mtok fast mode) plus a sandbox.network.strictAllowlist setting for locking down sandboxed command network access — release notes
• v2.1.218 moves /code-review to a background subagent so reviews no longer eat your active conversation context — release notes
• Opus 5 lands near Fable-level benchmarks at half Fable's price, a real inflection in cost-per-capability for agent workloads — AINews writeup

Coding agent tooling

• oh-my-pi (fork of Pi) ships hash-anchored edits, LSP/DAP integration, browser control and subagents in an ~80k-line Rust core, 60+ providers — repo
• addyosmani/agent-skills packages senior-engineer workflows (define→plan→build→verify→review→ship) as reusable skills for coding agents — repo
• Orca gives you an ADE to run Codex/Claude Code/OpenCode/Pi side-by-side in separate worktrees with unified usage tracking — repo
• bb is a self-hosted agentic IDE that orchestrates multiple coding agents across desktop/web/CLI/HTTP with live, steerable threads — repo
• openwiki (LangChain) auto-generates and maintains a linked Markdown wiki from your codebase, meant as agent-readable memory — repo
• archify turns a repo or system description into self-contained, interactive architecture/sequence/flow diagrams for agent skills like Claude Code/Cursor — repo
• Paritok claims 85% cost cuts and 3x longer coding-agent sessions via some context-compression trick — Product Hunt
• Prime Agent (Prime Intellect) pitches a coding agent that refines its own harness — Product Hunt

Small & local models

• Meta's Muse Glimmer: 30B open (Apache 2.0) model tuned for always-on local agent workflows, runs on a single consumer GPU — blog; community notes it still needs 32-64GB memory, limiting real consumer adoption
• Needle 2: 45M-param, 14MB tool-calling model that runs in 28MB RAM for phones/wearables/robots, but reportedly fails basic tasks (returned confidence=0 on a lock_door call) — details
• LiquidAI's LFM2.5 (2.6B) claims parity with models 4x its size, targeting reliability at tiny scale — model card
• GLM-5.2 marks a capability step-change for open agentic models per Interconnects' ongoing tracking — analysis
• h3.c brings native MiniMax-H3 inference to Apple Silicon via Metal, running on 64GB machines with quantization (needs workflow changes for tools like ComfyUI) — repo

Agent memory & context engineering

• TencentDB Agent Memory: team-level memory hub turning conversations/docs/code into four governed, shareable memory types (Chat Memory, Skill, LLM-Wiki, Code-Graph) — repo
• Reddit thread on a homegrown persistent-memory layer for Claude Code, addressing the "forgets everything between sessions" gap — thread
• Two weeks of notes cataloguing Claude failure modes invisible in any single turn, useful for long-session agent debugging — thread
• Someone's converting 10k+ new arXiv papers into Claude Code skills instead of re-researching from scratch each time — thread
• Mcptoon: MCP CLI client cutting tool-discovery overhead from ~10k tokens to ~350 by avoiding full schema dumps — repo; relevant as MCP tool counts climb past 100+
• Dan Luu's token-efficiency-by-language post resurfaces: dynamic/concise languages score lower token cost, but community pushback says the delta is small next to correctness — post

Agentic SaaS

• OmniRoute: free MIT AI gateway aggregating 290+ providers/500+ models behind one endpoint, with quota-aware fallback and RTK+Caveman compression claiming 15-95% token savings — repo
• Paperclip: open-source Node/React orchestration layer to run a "company" of AI agents with shared goal tracking and cost dashboards — repo
• GitHub ships a Copilot SDK for Java, letting enterprise Java devs drive Copilot via annotations and virtual threads — post