On July 3, Reuters reported that Meta is having trouble turning AI agents into reliable products.
That’s the whole story. It ran on a Friday, buried between OpenAI floating an equity stake to the US government and Google’s data centres pushing electricity use up 37%. Nobody stopped scrolling.
They should have.
Meta has the compute. Meta has the models. Meta has more engineers pointed at this problem than most countries have engineers, and a chip in production this September to make sure it never runs out of silicon. And the specific thing they cannot do — the thing that made the wire — is make agents reliable.
Now put that next to the other number everyone quoted this month. Gartner projects 40% of enterprise applications will have embedded agents by the end of 2026. Up from under 5% in 2025.
Both of these were reported as normal. They cannot both be normal.
Here’s the trick that makes 40% possible: nobody agrees what an agent is.
Anthropic drew the line clearly a while back, and it’s the most useful sentence in this whole discourse. Workflows are systems where models and tools get orchestrated through code paths you wrote in advance. Agents are systems where the model directs its own process and decides its own tool use.
That distinction sounds like pedantry until you sit through a vendor demo.
A “fully autonomous AI agent for sales operations” that turns out to be three prompt templates wired together with if/else statements is a workflow. There is nothing wrong with a workflow. Anthropic literally recommends them as the default. But it means that when a company tells you they’re deploying agents at scale, the honest follow-up is: how dynamic is the planning, actually? And at which step does the human stay in the loop?
Ask that question in enough rooms and the 40% starts to look less like a forecast and more like an accounting choice.
Because here’s what 40% adoption of “embedded agents” is going to look like in practice. A chatbot with a function call. A summariser with a database connection. A form-filler with retry logic. All of it shipped, all of it counted, most of it a workflow wearing an agent’s jacket.
Meta isn’t failing at that. Meta can build a form-filler in an afternoon. Meta is failing at the actual thing. And the actual thing is what everyone else is quietly not attempting while they book the credit for attempting it.
I spent the last few months building a multi-agent system for automated claims reimbursement. Multiple agents, a consensus mechanism where they argue with each other before committing to a decision, guardrails that can only ever tighten and never loosen, the whole thing wired into a real orchestrator with real queues.
It worked. It also taught me exactly what Meta is running into, and it isn’t the part anyone writes about.
The demo is easy. The demo is genuinely, embarrassingly easy. You can get a multi-agent pipeline doing something impressive in a weekend, and it will feel like magic, and you will want to post about it.
The reliability is where the entire mountain is. What happens on the ninth step when the model gets a partially-formatted response and confidently keeps going? What happens when two agents disagree and both are wrong? What do you write to the audit log when the decision emerged from a negotiation between three models and none of them can tell you why? Who signs off when the number that goes into someone’s bank account came out of a debate?
That last one isn’t rhetorical. It’s the question the July technology radars keep circling and nobody answers: has your organisation decided who is accountable when the agent makes a mistake? In production, errors aren’t hypotheses. They’re costs with a name attached.
The LangChain survey of 1,300+ practitioners this year has the shape of it in one pair of numbers. Around 89% of respondents have implemented observability for their agents. Only about 52% have implemented evals.
Read that again. Nine in ten teams can watch their agent work. Barely half can tell whether it worked.
That’s not an agent industry. That’s a very expensive dashboard industry.
And the top barrier to production in that same survey wasn’t cost — cost concerns actually dropped from the year before. It was quality. A third of respondents named it. The thing standing between agents and production isn’t money or compute or model IQ. It’s that the output isn’t good enough often enough, and nobody has a rigorous way to know when it isn’t.
While all of this is happening, the code underneath is drifting too. Greptile’s longitudinal data on pull requests reviewed through its platform found signs of end-to-end AI generation — bot authorship, co-author footers, agent branch prefixes — in 27.6% of merged PRs as of April 2026. In February 2025 that figure was 0.86%.
Separately, Sonar surveyed over 1,100 professional developers and found that only 48% say they always verify AI-generated code before committing it. Different datasets, different populations, so you can’t multiply them into a single scary ratio and I’m not going to pretend otherwise.
But you don’t need to multiply them. Just hold them next to each other and sit with it for a second.
We are building unreliable agents, on top of code that is increasingly written by agents, that fewer than half of us are checking.
Meta is struggling with the top layer of that stack. Meta.
This is the part where I’m supposed to say agents are a bubble. I don’t think they are, and I’d be a hypocrite — I build them.
Here’s what I actually think, having watched my own system succeed at the demo and then get hard exactly where Meta is getting hard.
The teams that win the next two years won’t be the ones with the most autonomous agent. They’ll be the ones who solved the boring things: orchestration, grounding, permissions, evaluation, security. Not model IQ. The stuff that doesn’t fit on a keynote slide because it looks like plumbing.
If you came into 2026 hoping for fully autonomous digital workers, adjust your timeline. If you came in hoping for software that can finally do multi-step work across real systems with a real audit trail — you got that. That’s the less braggable outcome and it’s by far the more interesting one. So when the 40% lands and the case studies start rolling out, ask the three questions that actually separate a system from a slide:
How dynamic is the planning, really? If a human wrote the sequence, it’s a workflow. Say workflow. It’s not an insult.
Where are the evals? Not the observability. Not the traces. The evals. If a team can watch it run but can’t tell you whether it ran correctly, they don’t have an agent in production. They have a very well-monitored guess.
Who’s accountable when it’s wrong? If nobody in the room has an answer, it isn’t in production. It’s in a demo that happens to have a URL.
Meta couldn’t answer these fast enough to ship. That’s not an indictment of agents.
It’s the price of the ticket, and most of the people claiming a seat haven’t paid it.
I build multi-agent systems and I’m still doing it after writing all of this. Tell me where I’m wrong — I’d rather find out in the comments than in a production incident.
Meta Can’t Make Agents Work. But Your Startup Can? was originally published in Stackademic on Medium, where people are continuing the conversation by highlighting and responding to this story.