Coding agents

• A week running a fully-autonomous Claude Code agent with its own co-signed wallet: real-world log of giving an agent an empty mandate and actual funds, worth reading for autonomy failure modes.
• workweave/router: drop-in proxy in front of Anthropic/OpenAI/Gemini that routes each prompt to the cheapest adequate model via an on-box embedder, claims 40-70% cost cut, direct hit on cost-aware routing.
• My AI agent suddenly started editing files from a completely different project: practical writeup on forcing an agent to stay scoped to its own repo.
• Your agent isn't confused, it's obeying a rule you stopped following three refactors ago: reframes "agent drift" as stale CLAUDE.md/rule rot after the codebase moved on.
• How to bring your software delivery workflow into GitHub with agent apps: GitHub ships four agent apps covering scope, secure, roll out, ship, all inside GitHub.
• We built a retrospective reader for Claude Code using Claude: tool that mines your own Claude Code session transcripts to generate retros.
• Domain-Driven Agents: argues for pairing .md domain docs with code entities so agents can navigate legacy, high-coupling codebases; community flags the post itself as possibly AI-written.
• Bit confused seeing the thoughts, but I'd rather see them vs Gemini hiding them: user take on extended-thinking transparency as a trust feature, not noise.
• Does Opus 5 verbosity affect its real world coding capacities compared to Fable: open question on whether Opus 5's chattier reasoning trades off against actual coding output quality.
• openJiuwen-ai/jiuwenswarm: multi-agent swarm framework wired into everyday messaging apps, with skill self-evolution and tool invocation.
• THU-MAIC/OpenMAIC: one-click multi-agent "classroom" generator, now with a steerable curriculum-planning agent on top of the original generator.
• The only thing I'll miss about Copilot: brief migration-away-from-Copilot gripe/nostalgia thread.

LLM mechanics & new models

• Introducing Hy4 Preview / Tencent release / ProductHunt listing: Tencent open-sources Hy4, 770B total / 49B active MoE params, 1M token context, 1.56TB weights on Hugging Face; community notes its ~5% cache pricing undercuts rivals' 10%+ and Hy3 already proved strong at agentic tasks.
• GLM-5.3: How Chinese labs keep stride with the frontier: Z.ai's new model shows a big benchmark jump, live now only in the coding plan, API and open weights (Hugging Face) coming in two weeks.
• Benchmarking Pocket-Scale Inference: Artificial Analysis + Liquid AI benchmark small models running directly on phones; community pushback is that it optimizes for flagship hardware while the real bottleneck is older phones/iPads capped at 4GB RAM.

Vibe-coding & repos

• tt-a1i/archify: agent skill for Cursor/Claude Code/Codex/OpenCode that compiles a typed JSON IR into self-contained, animated architecture/sequence/data-flow diagrams, no manual diagramming.
• I saw Tim's video on Codex vs Claude and I was amazed as a vibecoder: community reaction thread to a Codex-vs-Claude comparison, signal on vibe-coder sentiment shifting.

Agentic SaaS

• I built a simulator fleet so AI agents could actually drive iOS apps: infra for letting agents drive real iOS Simulators instead of mocked UI, relevant to agentic E2E testing.
• Osmantic/ODS: one-shot installer that turns a PC/Mac/Linux box into a private AI server (inference, chat UI, voice, agents, workflows, RAG, image gen), bundling what you'd otherwise wire together by hand.
• What image/video/music MCP server do you guys use?: community comparison thread on media-generation MCP servers (BudgetPixel, Higgsfield, OpenArt).
• Staats: lets your coding agent query site health/status directly, agent-facing observability play.
• cursor/plugins: official Cursor plugin spec plus a marketplace of ready-made integrations for dev tools and SaaS products.
• Cohere Parse 5: document parser turning complex docs, tables, and images into AI-ready structured data.