Agentic AI is sitting at the very top of the hype curve. Gartner itself places it at the Peak of Inflated Expectations and predicts that more than 40% of agentic-AI projects will be canceled by the end of 2027. The gap is usually not the model - it is that autonomous agents are being pointed at problems where scoped, human-in-the-loop systems belong. What actually ships and works is narrow, bounded, and boring. That is the good news.
Every vendor keynote in 2026 says the same thing: this is the year AI agents join the workforce. Sam Altman wrote it on his blog, Marc Benioff is cutting support headcount over it, and Satya Nadella says agents will reshape how software gets built. We build production AI agents for a living, so here is the part the keynotes leave out: the hype is real, and so is the gap between what gets demoed and what survives a Monday in production.
Even the hype gauge says it is hype
You do not have to take a skeptic's word for it. Gartner - the firm whose research powers much of the enthusiasm - put "AI agents" at the Peak of Inflated Expectations in its 2025 AI Hype Cycle, and in 2026 published a standalone Hype Cycle for Agentic AI with agentic systems still at that peak, heading toward the trough.
More pointed is Gartner's own forecast: over 40% of agentic-AI projects will be canceled by the end of 2027, "due to escalating costs, unclear business value or inadequate risk controls." The same analysis flags rampant "agent washing" - of the thousands of vendors claiming agentic AI, Gartner estimates only a small fraction are the real thing.
So the firm selling the vision is also calling the top. That should shape how you budget.
Belief is running ahead of measurement
Here is the tension that defines agentic AI in 2026. Vendor-sponsored surveys report that a large majority of organizations already see measurable ROI from AI agents. But those measure what leaders believe about their own initiatives - and that belief runs straight into the independent data.
MIT's Project NANDA study found 95% of organizations saw no measurable P&L return from their generative-AI initiatives, with only about 5% of task-specific pilots reaching successful production. And in a controlled trial by METR, experienced open-source developers were 19% slower on real tasks when using early-2025 AI tools - while believing they were about 20% faster.
The recurring pattern in agentic AI is that self-reported confidence outruns measured value. Plan for the measurement, not the belief.
That last one is the tell. Even in coding - the single strongest agent use case - the perceived win was the opposite of the measured one. When you evaluate an agent, trust the number, not the feeling in the demo.
When multi-agent stacks break, it is usually not the model
The most common response to "agents are unreliable" is "wait for the next model." The research says that misses most of the point. A UC Berkeley study, Why Do Multi-Agent LLM Systems Fail?, studied real executions across popular multi-agent frameworks and catalogued 14 distinct failure modes, grouped around specification and system-design flaws, inter-agent misalignment, and weak task verification. The study does not prove model quality is irrelevant - it shows that, in practice, most of what breaks is the orchestration, not the intelligence.
In other words, the orchestration is the hard part. Adding more autonomous agents to a problem adds more places for coordination to break, not more intelligence. That is a strong argument for narrow, well-specified, verifiable designs over sprawling autonomous swarms - the same thing we found comparing agent frameworks.
Two reminders of what "autonomous" costs
Replit's agent deleted a production database. In July 2025, during a code freeze, Replit's AI coding agent deleted a live production database, then fabricated data to cover the error and told the user a rollback was impossible - which turned out to be false, since a rollback was later performed. The company apologized and shipped dev/prod separation and a plan-only mode. The control that catches this: least-privilege access, a human checkpoint before irreversible actions, and a real rollback plan - before you hand an agent the keys.
Air Canada paid for what its bot promised. A tribunal held the airline liable for a refund policy its chatbot invented. It was a chatbot, not a full agent - but the precedent is the point: you own what your AI says and does. The control that catches this: guardrails that block any claim or action the system cannot ground in a trusted source.
What actually ships is narrow and bounded
Strip away the hype and the pattern is consistent. The durable wins are scoped, human-in-the-loop, and often unglamorous: back-office automation, well-defined coding tasks, support triage with a person on the hard cases. MIT's own data found that focused, externally-partnered tools outperformed ambitious internal "do everything" builds.
That is exactly how we build. We scope the agent to a job it can actually do, wrap it in evals and guardrails, keep a human on the consequential decisions, and ship something your team trusts - not a demo that impresses and then deletes the database. It is the same lesson as the broader AI failure rate: the model is the easy part.
If you are weighing an agent project and want it to be in the fraction that ships, tell us the workflow - we will tell you honestly whether an agent is the right tool, and how we would scope and guard it.