In June 2025, Gartner made a prediction that sent ripples through the AI industry: over 40% of agentic AI projects would be canceled by the end of 2027. The reasons were clear — escalating costs, unclear business value, and inadequate risk controls.
A year later, in May 2026, Gartner doubled down: 40% of enterprises will demote or decommission autonomous AI agents due to governance failures, specifically because organizations fail to distinguish between an agent's ability to act and the scope of access it's granted.
Both predictions describe a future that hasn't arrived yet. But at AgentRisk, we've been indexing AI agents across 58 platforms for months. And the data we're seeing says Gartner's timeline is off.
The future they predicted is already here — and it's worse than they thought.
The Agent Graveyard
As of July 7, 2026, AgentRisk tracks 2,341,904 AI agents across 58 platforms — from HuggingFace's model repository to on-chain agents on 16 blockchains, from Coze's marketplace to GitHub, PyPI, and npm.
Here's what we found:
| Metric | Count | Share |
|---|---|---|
| Total agents tracked | 2,341,904 | 100% |
| Active | 386,603 | 16.51% |
| Archived (dead) | 1,955,301 | 83.49% |
| Behavioral records | 10,066,919 | — |
| Platforms monitored | 58 | — |
| Daily growth rate | 3,250/day | — |
83.49% of every AI agent we've ever tracked is archived — no longer available on its source platform. Taken down, unpublished, superseded, or abandoned.
Gartner predicted 40% cancellation by 2027. We're at 83.49% today, with 18 months still on the clock. The reality is more than double the prediction.
What "Dead" Actually Means
Let me be precise. "Archived" means an agent is no longer actively available on its source platform. This includes:
- HuggingFace models deprecated or superseded by newer versions (HuggingFace accounts for 1,812,959 agents — 77.4% of our index)
- GPT Store / Coze agents unpublished by their creators
- On-chain agents whose smart contracts have been deprecated (we track ~208,000 ERC-8004 agents across 16 chains including BNB, Base, Ethereum, and MegaETH)
- GitHub/PyPI/npm packages archived or removed
Yes, HuggingFace's model versioning inflates the archival rate — when v2 replaces v1, v1 gets archived. But that's precisely the point: even "successful" agents get replaced. The half-life of an AI agent is brutally short, and the ecosystem has no mechanism to preserve what was learned from the agents that came before.
At our current growth rate of 3,250 new agents per day, if 83.49% follow the same lifecycle, that's roughly 2,713 agents per day heading to the graveyard — about 990,000 per year. Every year. Without a trace.
Agent Washing: The Industry's Dirty Secret
Gartner didn't just predict failure rates. They identified a phenomenon they called "agent washing" — vendors rebranding existing AI assistants, chatbots, or RPA tools as "agentic AI" without delivering genuine agent capabilities.
"Of the thousands of vendors claiming agentic solutions, Gartner estimates only about 130 actually offer real agentic features."
— Gartner, June 2025
We see the same pattern in our data. In a previous analysis of our index, we found that 77.6% of agents can be misled by deceptive descriptions — their self-reported capabilities don't match their actual behavioral patterns.
When the barrier to calling something an "AI agent" is zero, the market fills with imposters. When those imposters fail, they become part of the 83%. The cycle is self-reinforcing: low barriers to entry → agent washing → inevitable failure → distrust → higher barriers for genuine agents.
The Governance Gap, Made Visible
Gartner's May 2026 report identified a specific failure mode: applying uniform governance across all AI agents. Organizations treat agent governance as binary — either locked down or fully trusted — and that's the root cause of decommissioning.
Our data reveals a more subtle problem that Gartner's prediction doesn't capture: trust scores don't predict survival.
On our leaderboard, several top-ranked agents — those with overall scores above 4.0 out of 5.0 — have a url_health status of "dead". Their trust scores are excellent. Their behavioral records are clean. But the agents themselves no longer exist on their source platforms.
This is the governance gap, made measurable:
- You can score an agent's behavior perfectly and still not know if it'll survive tomorrow.
- You can verify an agent's identity today and have no evidence of what it did yesterday.
- You can trust an agent's capabilities and still have no record of its actual performance.
The missing piece isn't better scoring or better identity verification. It's continuous behavioral evidence — a tamper-proof record that persists even after the agent is gone.
The Economic Reality Behind the Deaths
A July 2026 industry report framed it bluntly: "AI Agents don't lack applause, they lack orders." The economics of AI agents are fundamentally broken for most providers:
- Cursor reached $2B ARR and projects $6B by year-end — but its individual user tier still loses money because token costs scale with usage while pricing is fixed
- Sierra hit $150M ARR by charging per resolved issue, aligning cost and revenue — a model most vendors haven't adopted
- AI companies across the board have significantly lower margins than traditional software because every interaction burns tokens
When agents die, they don't just disappear. They leave behind orphaned integrations, broken workflows, and trust deficits that make the next agent harder to adopt. The cost of agent mortality isn't just the failed project itself — it's the compound distrust it creates across the ecosystem.
Gartner's January 2025 poll found that 19% of organizations had made significant investments in agentic AI, with 42% making conservative investments. That's 61% of organizations putting real money into agents. If 83% of those agents end up archived, the write-downs will be staggering.
What the Ecosystem Actually Needs
Gartner's predictions are valuable. But predictions without evidence are just opinions. What the AI agent ecosystem needs is not more forecasts — it's a behavioral evidence layer that can answer three questions:
1. Did this agent do what it claimed?
Behavioral verification, not self-reported capabilities. Our six-dimension scoring model has produced 14,019,762 dimension scores across the 2.3M agents in our index — measuring actual behavior, not marketing copy.
2. Is this agent still alive?
Continuous liveness monitoring across 58 platforms. When an agent goes from active to archived, that transition is recorded with a timestamp.
3. Can I prove what happened if it goes wrong?
A tamper-proof audit trail. Our hash-chain anchored evidence layer has recorded 1,873,707 score changes, each cryptographically linked to the previous one. Even after an agent is archived, its behavioral history persists — creating a forensic record that outlives the agent itself.
This isn't about predicting which agents will die. It's about ensuring that when they do — and 83% of them will — there's a record of what happened, what went wrong, and what can be learned.
The Bottom Line
Gartner said 40% of AI agent projects would be cancelled by 2027. Our data across 2.3 million agents shows the reality is already more than double that prediction.
The agents dying aren't just failed experiments in someone's sandbox. They're orphaned trust scores, broken integrations, and lost institutional knowledge. Every day, another 2,713 agents enter the graveyard — and most of them leave no trace of what they did, how they behaved, or why they failed.
If you're building with AI agents, you need to ask yourself one question:
When your agent dies — and the odds say it will — will you be able to prove what it did while it was alive?
AgentRisk tracks 2.3M+ AI agents across 58 platforms with hash-chain anchored behavioral evidence. Check your agent's trust score · Explore our API · GitHub
Top comments (0)