In January 2026, an Austrian developer published an open-source AI agent framework called OpenClaw. Within 72 hours it had 60,000 GitHub stars. By March it had surpassed React to become the most-starred project in GitHub's history. NVIDIA CEO Jensen Huang called it "probably the most important software ever released."
When I read that, my first thought as a PM was: this isn't just a developer story.
From Suggestions to Actions
Until recently, AI tools operated in a single mode: recommend. "Here's a draft. Here's a code suggestion. Here's a bug you might want to fix." You still had to decide, click, execute.
OpenClaw and this generation of AI agents work differently. Given the right permissions, they complete the task. They open the email, check the calendar, create the file, fill the form. You define the goal; the agent handles the steps.
That distinction sounds subtle. For a product manager, it's enormous.
Think about how most "AI-powered" features work today: they surface information or draft content, then wait for the user to act. That model still places the cognitive burden on the user. An agent removes that burden entirely. The product doesn't help you do the thing — it does the thing.
What This Looks Like in a Real Product
I manage four health SaaS products: DentalBulut, Medibulut KYS, Diyetbulut, and medibulut.com. Our users are clinicians — doctors, dentists, dietitians — who see patients from 9am to 8pm and have almost no time to look at a screen.
For these users, a "suggesting" AI was never going to move the needle. A doctor mid-consultation isn't going to type "AI, reschedule this appointment." But an agent could automatically shift that appointment after the consultation ends, send the patient an SMS, and add a note to the patient record — with a single approval tap.
That's the difference between a feature that's technically impressive and one that actually changes a user's day.
When I look at our backlog now, I apply a new filter: is this feature suggesting or doing? Suggesting features still have value. But doing features are the ones that create new product categories. Every product slaps an "AI-powered" badge on a suggestion engine these days. The real differentiation is whether your product takes action.
The Hard Part: Getting Autonomy Right
OpenClaw going viral proved one thing clearly: users are ready for agents. The infrastructure isn't.
NVIDIA recognized the gap immediately and launched NemoClaw — essentially a governance and safety layer for running OpenClaw in enterprise environments. Because giving an agent "full system access" is also a data breach waiting to happen.
As a PM, this is where my job gets genuinely difficult. Designing agent features isn't just about capability — it's about calibrating the right level of autonomy for your specific users and context.
In healthcare, this is especially sensitive. A miscalibrated agent that acts on incomplete patient data isn't just a UX failure — it's a regulatory risk and a trust problem that can take years to recover from.
I've started thinking about autonomy in three levels:
Level 1 — Autocomplete: The user initiates, AI finishes. (Drafting a prescription template.) Low risk, moderate value.
Level 2 — Trigger-based automation: AI acts when a condition is met. (Send SMS 24 hours before an appointment.) We already do this. High value, well-understood.
Level 3 — Proactive agent: AI initiates, user approves. (Detect a follow-up gap and start the outreach flow.) In healthcare, this level is early — but it belongs on the roadmap.
Three Things I'm Doing Differently Now
The OpenClaw story isn't a distant technology trend. It's a signal that the baseline is shifting. Here's how it's changing my day-to-day as a PM:
First, I'm auditing our existing workflows with one question: could an agent handle this? Every manual step in a user's workflow is a potential agent opportunity. Most of them were invisible to me before.
Second, I've added a new question to user research: how much control are you willing to give up? This sounds simple but it surfaces wildly different answers across user segments. A clinic owner running three locations thinks about this very differently than a solo practitioner.
Third, I'm designing safety and approval mechanisms upfront rather than retrofitting them. In the agent era, "undo" and "preview" aren't nice-to-haves — they're table stakes. Building trust with autonomous features requires making the agent's reasoning visible before it acts.
The agent era isn't coming. It's here. And the PMs who start treating "does the product act?" as a core product question will be building something fundamentally more useful than everyone still optimizing their suggestion engines.
What's your take? Have you shipped an "AI does it" feature yet, or are you still in suggestion mode?