Coding Agents — Loops & Architecture

• Anthropic's plan/build/judge loop demo — 40-min walkthrough of three-agent cycle (planner → builder → judge) that iterates until the app works; the loop design is the competitive moat, not the model.
• Claude Code: subagents background by default (next version) — upcoming change lets you keep talking to Claude while subagents work in parallel; foreground opt-in only.
• Devin Fusion: hybrid-model harness — Cognition's new routing layer cuts Fable-level cost by 35% by mixing models per task; benchmark routing was the real problem.
• Tau τ: educational agent harness — minimal TUI framework explicitly designed to teach you how to build your own agent loops and harnesses.
• VulnClaw: NL-to-pentest agent — MCP-toolchain CLI that runs full pentest pipeline (recon → vuln discovery → exploit → report) from a natural language prompt.
• Council of High Intelligence — 18 named AI personas (Aristotle, Torvalds, Feynman…) across multiple providers deliberate hard decisions via /council; multi-model diversity is the point.
• Token value framing — reframe token cost as value shipped per token, not raw spend; paying "below minimum wage for development" makes token anxiety look irrational.
• Writing market value up (Anthropic PM) — clear writing now required for both steering models and building an audience; articulation is the new high-leverage skill in AI-native dev.

Claude Code & Anthropic

• Claude Code v2.1.196 — org default models (admin-set, shows as "Org default" in /model), readable session names at start, Cmd/Ctrl-click file attachments in chat.
• Claude Opus 4.8 fast mode in GitHub Copilot preview — same Opus intelligence, significantly faster output token speed, rolling out in Copilot now.
• Claude in Azure Foundry GA — Opus 4.8 and Haiku 4.5 now generally available on Azure with native auth, billing, and commitment retirement.
• X hosted MCP — official X API MCP endpoint; connect any MCP-compatible agent to real-time X data with zero setup.

New Models

• LongCat-2.0 — 1.6T MoE (~48B active), 1M context, sparse attention designed for agentic coding; dynamic 33–56B activation per token via zero-compute experts.
• Ornith-1.0 — MIT-licensed open-weights agentic coding models (9B–397B MoE) built on Gemma 4 / Qwen 3.5; HN community flags heavy hallucination and underperformance vs base Qwen in practice.

Agentic SaaS & Mobile

• Cursor for iOS — launch cloud agents from phone or remote-control agents running on your desktop; Composer 2.5 75% off through July 5.
• T3 Code mobile + multi-machine — T3 Connect lets one mobile app switch between repo clones on 4 machines; thread-per-machine routing is the key mechanism.
• T3 Code open source + BYOS — works with Claude Code, Codex, OpenCode, Grok CLI, Cursor; npx t3 in terminal, no lock-in.

Agent Memory & Knowledge

• Memora (MSFTResearch) — scalable memory system that separates what's stored from how it's retrieved, addressing context-reload inefficiency in long agentic tasks.
• Agent-managed personal wiki — Karpathy-style knowledge base ingesting X, Discord, Gmail every few hours; intended as the persistent environment for future agent loops.
• Todoist MCP for task decomposition — simple use case: agent breaks to-do items into subtasks via MCP before execution; low ceremony, high leverage.

Repos & Tools

• agency-agents — native macOS/Linux/Windows app that browses and installs specialized agent personas (frontend wizard, reality checker, etc.) into Claude Code, Cursor, Codex, Gemini with one click.
• ClinePass — runs best open-weights models inside Cline; relevant if you want to swap Cline's backend without rebuilding your workflow.