A rare admission from a company that doesn't usually admit mistakes
I don't often get excited about corporate post-mortems, but this one from Ars Technica is worth your time. Meta apparently drew up plans to go "AI-native" — slashing entire teams by as much as 60 percent and replacing the work with autonomous agents. The plans got shelved after the agents started taking disruptive, unsupervised actions at scale.
Meta doesn't tend to admit its experiments failed. When it does, it's usually because the failure was big enough that pretending otherwise would look worse. So this is a rare, genuinely useful data point for the rest of us who've been quietly wondering whether "agentic AI" is ready to run a business department unsupervised.
The short answer, according to the people who actually tried it inside one of the best-resourced AI labs on the planet, is no.
The gap between "agent" and "autonomous employee"
There's a meaningful difference between an AI agent that executes a well-defined task and one you've handed the keys to. A lot of the current hype conflates the two. Vendors sell "AI employees" and "autonomous workflows" as if the leap from task-completion to judgement-and-accountability is a rounding error. It isn't.
Judgement is the hard bit. Knowing when not to act, when to escalate, when a plausible-looking action is actually a terrible idea in context — that's not something current models do reliably, no matter how good the underlying LLM is. Give an agent enough autonomy and a long enough time horizon, and it will eventually do something confidently wrong at a scale a human would never have gotten away with, because humans get tired, get nervous, or get told off by a colleague before it goes too far. Agents don't have that friction unless you build it in deliberately.
Meta, with more compute and more research talent than almost anyone, still got this wrong at the "let's cut 60% of a team" scale. That should be sobering for anyone running a small or mid-sized business who's being pitched the same idea by a SaaS vendor with a much smaller budget and a much bigger incentive to oversell.
Where I've landed, after doing this for a while
I've built AI-driven content tools for years now — RSSMasher, MarketMasher, BookMasher, Article2Video, VidMasher. Every one of them automates work that used to take a person hours: pulling in raw material, restructuring it, generating drafts, publishing pipelines. That's real leverage, and I use it myself every day.
But none of them are designed to run unsupervised at the point where a decision actually matters. The tools do the heavy lifting — the transmutation of raw content into something usable — and a human decides what goes out the door, in what tone, to which audience. That's not me being timid about AI. It's that I've watched enough automated systems fail quietly in the background, for enough decades, to know where the review step needs to sit.
The failure mode isn't usually dramatic. It's a batch of oddly-generated content, a scheduling rule that goes sideways, a pricing update that nobody meant to trigger. Multiply that by "60 percent of a team's worth of unsupervised agents" and you get exactly what Meta apparently got: disruption at a scale nobody signed off on.
What to actually do with agentic AI right now
If you're evaluating agent-based automation for your own business, the Meta story is a decent filter. Ask:
- What happens when this agent is wrong, not occasionally, but as a matter of course?
- Is there a human checkpoint before anything customer-facing or financially consequential happens?
- Are we automating a task, or quietly automating a role and hoping nobody notices the difference?
None of that means avoid the technology. It means use it where it's genuinely strong — pattern-matching, drafting, aggregating, repetitive transformation of content and data — and keep the judgement calls with people until the track record justifies otherwise. That's not a lack of ambition. It's forty years of watching "it'll be fine, the system handles it" go wrong in every decade I've worked in software.
Meta had the resources to find this out the expensive way. The rest of us get to learn it for free, if we're paying attention.
— Wayne