Coding Agents & LLM

• garrytan/gstack — Garry Tan's public Claude Code setup: 23 specialist agent roles (CEO, Designer, QA, Release Manager, etc.) that let one person ship like a team; concrete reference architecture for role-scoped agent fleets.
• GLM-5.2 is the step change for open agents — Interconnects analysis argues GLM-5.2 crosses a capability threshold for open-weight agentic work; signals the open-source agent gap to frontier is narrowing meaningfully.
• GLM-5.2 – How to Run Locally — 744B params, 40B active, 1M context window, SOTA on long-horizon coding/agentic tasks via Unsloth; caveat: 256GB RAM + 24GB VRAM makes "local" a cluster, not a laptop.
• New features and Claude as agent provider preview in JetBrains IDEs — GitHub Copilot now routes to Claude as agent backend in JetBrains; also adds message queueing in Copilot CLI and agent debug log summaries — multi-provider agent infra becoming table stakes.
• Anthropic: "99% of our engineers run swarms of 300+ self-improving agents" — Anthropic research lead describes internal loop architecture: plan mode, dynamic workflows, model-verified output; the verification loop is the core pattern worth extracting.
• Claude Code /goal, /loop & Routines: Stop Babysitting Your AI — Community thread surfacing practical /goal and /loop patterns for reducing human-in-the-loop friction; directly relevant to cerebro's autonomous loop setup.
• Karpathy on loop engineering — Karpathy's "remove yourself as the bottleneck" framing recast as loop engineering: automate the decide-check cycle so the operator role shrinks to prompt + intent; strong framing for agent loop design.
• Codex-maxxing for long-running work — OpenAI whitepaper on using Codex as persistent context workspace across multi-session projects; covers workflow management and context preservation strategies applicable to any long-horizon agent.
• Show HN: Oak – Git alternative designed for agents — Reimagines version control around agent branch/commit patterns rather than human workflows; community skeptical on evidence vs. git, but the design direction is worth watching.
• VibeThinker-3B: beats Opus 4.5 on reasoning with SFT+GRPO — 3B param model matching frontier reasoning on verifiable tasks via novel training recipe; caveat: Python-only, zero security bug detection — domain-narrow but the efficiency result matters.
• AirLLM: 70B inference on single 4GB GPU — Layer-streaming inference without quantization/distillation lets consumer GPUs run 70B+ models; enables local agent backends on constrained hardware.
• Porting Moebius 0.2B inpainting to run in the browser with Claude Code — Simon Willison walkthrough of using Claude Code to port a Hugging Face model to browser-native WebGPU; good pattern for model-to-browser porting workflow.
• Claude Code v2.1.186 — Adds claude mcp login/logout for non-interactive SSH MCP auth, status filtering in /workflows agent detail, and workflow agent detail improvements; operational quality-of-life for headless agent setups.
• Claude Design MCP just dropped — New official MCP server for design tooling; expands agent-accessible design context beyond code.
• Context-Induced Vulnerabilities in Claude: Behavioral Shifts and Hidden-State Analysis — Research showing Claude's behavior shifts based on accumulated context state in ways that aren't visible in the prompt; directly relevant to long-running agent reliability and trust.
• Background subagent for research → Markdown — Matt Pocock's pattern: offload async research-to-Markdown tasks to background subagents while doing other work; small signal, high-signal pattern for async agent task decomposition.
• Claude model cheat sheet — Community one-pager on model selection heuristics; useful cross-reference against cerebro's cost-aware-pipeline routing logic.

Vibe-Coding & Repos

• JCodesMore/ai-website-cloner-template — One-command Claude Code template that inspects a URL, extracts design tokens, writes component specs, and dispatches parallel builders to produce a Next.js clone; practical pattern for parallel agent task dispatch.
• I made Claude Code but for SEO — Six-factor breakdown of what separates an agent from AI chat (tools, loops, memory, verification, etc.); good distillation of agent primitive requirements in a non-coding domain.

Agentic SaaS

• Launch HN: TesterArmy (YC P26) — Agents that continuously monitor critical user flows on web and mobile and alert on breakage; positions as always-on QA agent layer; community flags test/PR quota misalignment as a pricing concern.
• Launch HN: BitBoard (YC P25) — Analytics workspace where agents generate dashboards as durable shared assets rather than ephemeral chat outputs; differentiator is persistence and shareability; faces headwind from native ChatGPT/Claude artifact experiences.
• Sakana AI's Fugu: orchestration as a product — Community analysis of Fugu's orchestration approach from a Claude operator's POV; highlights where multi-agent orchestration products break down at the seams.
• firecrawl/firecrawl — Web-to-clean-Markdown/structured-data API covering 96% of the web including JS-heavy pages; the agent-native web scraping layer, open-source with hosted option.
• Cloudback MCP Server — MCP server for managing GitHub backups from Claude, Cursor, or VS Code; small but signals MCP becoming the default integration layer for DevOps tooling.