Coding agents & harnesses

• OneCLI (YC S26) launches OSS sandboxed agent harness for teams, one agent per employee with scoped access; HN pushback wants method/path/field-level approval grants, not endpoint-level, to stop confused-deputy attacks.
• fx ships a 6.39MB Zig-native coding agent CLI, betting minimalism/embeddability wins over feature-complete harnesses in a crowded field.
• AGENTS.md feature request asks Claude Code to support the emerging cross-tool AGENTS.md standard (Codex/Amp/Cursor already do) instead of CLAUDE.md lock-in; thread notes symlink already solves it technically.
• Claude Managed Agents vs open source head-to-head on same 14 tasks, same model, worth a read if picking managed vs self-hosted agent infra.
• Keeping a coding agent consistent across sessions crowdsources what actually holds context/behavior stable session to session.
• Autonomous verify/review/repair tip argues for self-checking loops as default agent workflow, not human-gated review only.
• Multi-agent orchestration visualized shows a concrete orchestration topology in the wild.
• /implement skill with full dev flow is a homegrown implement-skill pattern, comparable to this session's own /implement.
• Ornith-1.5 extends self-scaffolding into a full self-improvement RL loop (model proposes tasks, builds scaffolds, trains on own rollouts); 35B runs well on consumer hardware but open-weights status unclear.
• qwen 3.8 27B agent harness poll surveys which local harnesses pair with that model for coding.
• Claude Code + Codex on one keyboard is a hardware hack, RGB F-row lane + summon key per agent session.
• Claude reconstructs WordPress plugin deletion history via custom connector, notable case of agent forensic reasoning over legacy infra.
• Custom MCP connectors broken 1+ week, Anthropic support auto-replied without reading report, worth tracking if you rely on custom MCP connectors.
• Structuring an AI-assisted React Native rewrite crowdsources workflow for large agent-driven mobile rewrites.
• Benchmarking GLM-5.3 for ordinary coding asks for practical eval methodology, not just leaderboard scores.
• Long-context sessions after llama.cpp restart, context-persistence pain point for local long-context coding sessions.

LLM mechanics

• Emulating ALiBi with RoPE walks through reproducing ALiBi's linear positional bias inside RoPE's rotation scheme, relevant if you care about KV-cache/positional-encoding tradeoffs.
• 48 AI-generated web apps clicked in real browser finds the pricier model failed more than the cheap one, real signal against blind model-tier scaling for codegen.

Agentic SaaS

• SnapState pitches persistent state for AI agent workflows, i.e. checkpointing agent runs across restarts.
• DevCake is a self-hosted OSS "software factory" for local-model agent pipelines.
• Hermes companion app lets you chat with your self-hosted agent from phone, fully open source.

Vibe coding

• How to vibecode without being an idiot is a discipline/guardrails thread for AI-assisted coding workflows.
• Imba, 7-years-in-the-making web language with dimension types and native web components; interesting but bakes framework opinions (memoized DOM) into the language itself, limiting reach beyond web.

CLI/TUI

• Redbean, single-file web server executable that embeds a static site as both zip content and a binary running unmodified across six OSes, genuinely novel packaging trick.

Extensible software

• Extensible Software in the age of LLMs argues LLM codegen commoditizes long-tail tool-building (30-min DIY beats stale pro tools), but mass-market users still weigh reliability over customization.