Coding agents

• GLM-5.2 is the step change for open agents — open-weight model closes gap with closed agent models on terminal/agentic tasks, worth watching if picking non-Anthropic backends.
• Claude Code v2.1.239 — cost estimates now bake in 1.1x US-only-inference premium for data-residency workspaces, fullscreen renderer now defaults on Bedrock/Vertex/Foundry installs.
• Claude Fable 5 and new AI safety fables — Anthropic's new model release paired with safety-classifier changes to research-query handling, relevant if routing through Fable.
• fx by Vercel — tiny open-source coding agent from Vercel, minimal-footprint entrant in crowded agent-harness space.
• OneCLI (YC S26) — OSS sandboxed agent harness giving each employee a scoped sandboxed agent; crowded field, $499/mo pricing hard to justify without fine-grained policy controls below endpoint level.
• llmfit — CLI benchmarks which local models actually run at usable tok/s on your hardware, crowdsources real numbers via TUI, no cloud account needed.
• Unpacking ChatGPT Work: the Agent for a Billion Users — deep dive on OpenAI's agent memory/deployment architecture at consumer scale.
• One long Claude Code thread vs parallel sessions — community thread on when single long-context sessions beat parallel worktree sessions and where each breaks down.
• Router by Ramp — token-cost routing layer pitched as saving money on LLM spend.

CLI/TUI

• llm-openrouter 0.7 — Simon Willison's OpenRouter plugin for llm now surfaces reasoning traces, compatible with LLM 0.32.
• TUI that interviews you on missing gaps in your project prompts — community-built prompt-gap-finder TUI, similar territory to spec-interview tooling.

Vibe-coding & agent frameworks

• obra/superpowers — composable skills framework plus development methodology for coding agents, structured alternative to ad-hoc prompting.
• mattpocock/skills — Matt Pocock's daily-driver agent skills, small/composable/model-agnostic, positioned against heavier process frameworks (GSD, BMAD, Spec-Kit).
• apache/maka — Apache Incubator project: local-first agent workspace logging every model message, tool call, and permission decision as an append-only recoverable log.
• affaan-m/ECC — agent-harness perf-optimization system (skills, instincts, memory, security) spanning Claude Code, Codex, Opencode, Cursor.
• ruvnet/ruflo — meta-harness adding 100+ specialized agents, swarm coordination, and self-learning memory on top of Claude Code/Codex.
• SnapState — persistent state layer for AI agent workflows, Cloudflare-gated signup.
• Plow Latch — scoped-access sandbox for running AI agents on macOS.

Agentic SaaS

• GitHub Copilot in Slack — Copilot CLI/app agentic capabilities now reachable via @GitHub mentions in Slack, public preview.
• GitHub Copilot shared work in Microsoft Teams — turns a Teams thread into a collaborative agent session others can watch/direct.
• Antigravity IDE Extensions — Google's Antigravity agents now embed inside existing editors instead of a standalone IDE.
• Epho — cloud runner for Claude Code, Codex, or Opencode against your repo.
• OneCLI — same OneCLI harness (see above), ProductHunt listing.
• Actx0 — memory infrastructure product for AI agents.
• OzBrain — shared cross-agent knowledge brain for Claude/ChatGPT/Cursor/team; LLMs still weak at summarizing large non-trivial text batches, core risk for this category.