October 16, 2026
Replit CEO Amjad Masad told TechCrunch he expects the future of work to look like a "self-driving company" -- AI agents handling the day-to-day grind while humans set direction and make the calls that matter. It's a catchy phrase, and it's also exactly the pitch Replit is building toward with its agent-heavy coding tools. But set the branding aside and the underlying idea is worth taking seriously: more and more of the repetitive work inside a business -- data entry, status updates, routing tickets, chasing invoices -- is becoming something software can just do, not something a person has to babysit. The honest caveat is that 'self-driving' implies a level of trust and reliability most agent systems haven't earned yet. Anyone who's watched an AI agent confidently misfile a record or loop on a task knows the gap between demo and dependable. The businesses that benefit from this shift won't be the ones that hand over the keys blindly -- they'll be the ones that pick where automation actually removes grunt work without removing oversight, which is the same philosophy behind workflow automation done right: automate the task, not the judgment.
Meta launched Muse Spark Open, an open-weight model built to run natively on Apple silicon, in partnership with Apple. That's a notable pairing: Apple has been cautious about outside AI partnerships, and Meta has leaned hard into open weights as its competitive wedge against closed labs. Put together, it means a business could run a capable model entirely on a Mac, with no data leaving the building and no per-token bill arriving at the end of the month. For companies that have been nervous about sending customer data to a third-party API, that's a genuine unlock, not just a cost play. It also raises the bar on what 'local AI' can mean for small and mid-size teams who don't have a GPU cluster, only laptops. The tradeoff is real too: on-device models still generally trail the frontier closed models on raw capability, and 'runs on a Mac' is a narrower promise than 'runs everywhere.' Still, for teams already weighing buy vs. build vs. ViibeStack decisions around their software stack, this is one more sign that the cost and control math on AI is shifting in the buyer's favor, not the vendor's.
Anthropic is giving researchers at METR employee-level access to look at its risks, incidents, and internal practices, and -- this is the part that matters -- Anthropic has given up the right to block what METR publishes. Most AI safety oversight so far has been self-reported or voluntary, which means a company can always quietly edit the parts that look bad. Removing the veto is a small, specific decision, but it's the kind of credibility signal that's rare in this industry: an outside party gets to say the uncomfortable thing and the company can't stop it. For businesses evaluating which AI vendors to actually trust with sensitive data and workflows, that distinction between 'we audit ourselves' and 'we let someone else audit us and publish it anyway' should matter a lot more than it currently does in most procurement conversations. It's the same instinct behind why we built out a real Trust Center rather than just a page of marketing claims -- oversight only means something if the company being watched can't edit the verdict.
If your company had to choose today, would you rather hand repetitive work to an AI agent, run a capable model locally on your own hardware, or trust a vendor's safety claims only once an outside party can publish the bad news too?
Sources