SoftBank just dropped $40 billion into OpenAI. Not for chatbots. For agents.
I've had three of them running my business for the past month. Here's what they're actually doing, what broke, and whether that $40B bet makes sense from someone who built their own agents from scratch.
The 3 agents and what they actually do
The first is a content engine. It hooks into my workspace and automatically logs everything I debug or build into a content bank. No more "what should I film today?" I query the bank, it surfaces the best ideas ranked by relevance and trending topics. It also runs a weekly scrape of top performing AI content on YouTube and TikTok. I see what's winning in my niche and reverse engineer my own angle.
The second is an email agent with a 4 tier classification system: (1) urgent/reply needed, (2) brand deals to surface, (3) newsletters to archive, (4) spam to kill. The critical layer: it also scans the spam folder daily. Gmail's own classifier was burying legitimate brand deal emails $1,000 to $3,000 inquiries sitting unread. The spam rescue layer alone has probably paid for this entire system multiple times over.
The third is a brand deal hunter (still in build phase). It researches AI companies with recent VC backing, scores them for fit, and drafts outreach emails. Not live yet but the infrastructure is wired.
The technical stack underneath
Everything runs through natural language pipelines, not code. I use Claude (Anthropic) as the primary reasoning layer. The agents talk to each other through a shared memory system MEMORY.md as long term state, daily markdown files for session memory. All state is managed in JSON files and triggered via cron jobs (43+ scheduled tasks running on macOS launchd + a Hetzner VM).
The hard gate system is critical: every action that touches the outside world (email sends, social posts, anything public) goes through an approval layer before it fires. This was built the hard way after the email agent sent a test email during what was supposed to be a dry run.
What the $40B bet is actually saying
AI crossed a threshold. It's not a chatbot anymore. It can execute. Email pipelines that close deals. Scraping systems that find leads. Content machines that run themselves. The gap between people who understand this and people who don't is growing every quarter.
The businesses building agent systems now are going to be operating at a fundamentally different level than the ones who figure it out in 2027.
What's the first thing you'd automate with an agent? Drop it below.
Top comments (0)