Agents meet the office
The agents clocked in this week: an Office suite built for them is the fastest-growing repo on GitHub, Tencent open-sourced the sandbox layer most agent frameworks skip, and OpenAI shipped a plugin that puts Codex inside Claude Code — multi-model workflows going mainstream. Off to the side, Apple sued OpenAI over hardware secrets and Google quietly made Gemini 3.6 Flash the new default workhorse. Busy week for a technology that supposedly isn't reliable yet.
The tools
OfficeCLI productivity open source
Office suite built for AI agents. Read, edit, and automate Word, Excel, and PowerPoint from any agent. No Office installation required.
7K new stars this week because every AI agent builder has hit the Office automation wall. Single binary, free, works everywhere.
CubeSandbox agents open source
Instant concurrent sandbox for AI agents -- lightweight, secure, purpose-built for agent code execution.
10K stars and growing. Sandboxed agent execution is the missing piece most agent frameworks skip.
system_prompts_leaks research open source
Extracted system prompts from Claude, GPT-5.6, Gemini 3.5, Grok, Cursor, Copilot, and more. 56K stars, 7K this week.
Fascinating peek under the hood — seeing what instructions major AI companies bake into their models tells you more about their priorities than any press release.
Codex Plugin for Claude Code coding open source
OpenAI plugin that lets you use Codex from inside Claude Code for code review and task delegation.
28K stars. The multi-model agent workflow is becoming the default pattern for serious developers.
Worth reading
Refreshing take: an agent is just a loop. 40 lines of Python, no LangChain, no vector DB. The frameworks solve problems you do not have yet.
Measured breakdown of the Entelligence AI leak claiming Gemini 3.5 Pro beats both Fable 5 and GPT-5.6 internally. Good community skepticism calibration.
The vibe coding hangover is real -- every minor error costs expanding context windows and massive token bills. Smart teams are moving to structured agentic workflows instead.
From the big labs
Apple alleges OpenAI told candidates to bring actual hardware prototypes to interviews. Complaint calls OpenAI hardware division rotten to its core. This is going to get extremely ugly.
Even the biggest AI spender on earth cannot get internal agents to work reliably yet. This is the honest state of agentic AI in mid-2026 and it should temper expectations.
Cheaper output ($7.50/M, down from $9), roughly 17% fewer output tokens, better coding and computer-use scores — and a Gemini 4 tease in the same breath. The workhorse tier is where production traffic actually lives.