AI Agent Adoption Statistics: The Numbers, Including Ours
Every AI agent statistics roundup cites the same enterprise forecasts. Almost none measures what individual developers are doing — so we did, and the gap between the two stories is the interesting part.
The enterprise numbers
Gartner forecasts that 40% of enterprise applications will feature task-specific AI agents by 2026, up from less than 5% in 2025 — an eightfold increase inside a year.
In its best-case scenario, Gartner projects agentic AI driving roughly 30% of enterprise application software revenue by 2035, surpassing $450 billion, up from 2% in 2025.
And the number that belongs in the same paragraph rather than a different article:
Gartner also warns that over 40% of agentic AI projects are at risk of cancellation by 2027.
Both of those are Gartner’s. Read together they describe a technology being embedded rapidly and abandoned frequently — which is a more useful picture than either figure alone, and it’s why we’d treat any single adoption percentage with caution.
What we measured: where the attention actually is
Enterprise forecasts describe procurement. GitHub describes what individuals choose. We queried the GitHub API on 19 August 2026 for eight self-hosted agent and automation projects:
| Project | Stars | Created | Age (months) | Stars/month |
|---|---|---|---|---|
| OpenClaw | 386,511 | 2025-11-24 | 8.8 | 43,901 |
| Hermes Agent | 231,778 | 2025-07-22 | 12.9 | 17,952 |
| n8n | 200,944 | 2019-06-22 | 85.9 | 2,339 |
| Dify | 152,679 | 2023-04-12 | 40.2 | 3,794 |
| Open WebUI | 148,990 | 2023-10-06 | 34.4 | 4,328 |
| OpenHands | 84,273 | 2024-03-13 | 29.2 | 2,886 |
| Activepieces | 23,842 | 2022-12-03 | 44.5 | 536 |
| Agent Zero | 18,893 | 2024-06-10 | 26.3 | 719 |
Ranked by velocity, the picture is nothing like the star totals:
- OpenClaw — 43,901/month
- Hermes Agent — 17,952/month
- Open WebUI — 4,328/month
- Dify — 3,794/month
- OpenHands — 2,886/month
- n8n — 2,339/month
OpenClaw is accumulating stars roughly 19 times faster than n8n, and it did not exist before late November 2025. Hermes Agent, created in July 2025, has already passed n8n’s total.
Two things worth drawing out:
The personal-agent category is younger and moving faster than the automation category. n8n has been compounding since 2019 and sits sixth by velocity. The two projects at the top are both under fourteen months old.
Star velocity measures attention, not use. A star is a bookmark. It costs nothing, it doesn’t imply installation, and a project trending on Hacker News collects them from people who will never run it. What this table reliably shows is where developer interest is concentrated right now — which is a leading indicator of ecosystem, documentation and available help, and that genuinely affects whether self-hosting one of these works out.
We later put a number on how misleading a single metric can be. Ranking nine self-hosted projects by Docker pulls and by stars produced almost inverted lists — a rank correlation of −0.70, with Nextcloud first on one measure and last on the other. Stars capture arrival; pulls capture entrenchment.
What it costs to actually run one
The number readers come for and no analyst report publishes. From our own pricing research across this series:
| Component | Monthly |
|---|---|
| VPS, 4GB — the floor for a complete agent | $6.49–14.99 |
| VPS, 8GB — once browser automation is on | $8.79–14.99 |
| Model API, typical personal use | $2–15 |
| Model API, frontier model with unmodified heartbeat | up to ~$85 |
| Typical all-in | $14–27 |
The server is the smaller half of the bill, which is the single most common misconception in this category. And the variance sits almost entirely in the API line: the difference between $2 and $85 is a heartbeat interval and a model choice, not your hosting.
Two specific figures that recur across our testing:
4GB is the floor and 8GB is the honest number once a browser is involved, because a headless Chromium instance wants 1–2GB entirely to itself. That single fact accounts for most reports of agents dying without explanation.
Almost every self-hosted agent wants the same 4GB. We compared the published requirements for OpenClaw, Hermes Agent, Agent Zero, OpenHands, Dify and n8n — the floors converge, so the spec sheet doesn’t differentiate them.
The security numbers
Adoption statistics rarely include the failure statistics, and in this category they’re substantial.
At the disclosure of CVE-2026-25253 — a one-click remote code execution flaw in OpenClaw rated CVSS 8.8 — over 40,000 instances were exposed on the internet and roughly 63% were assessed as vulnerable.
The supply chain followed: the ClawHavoc campaign distributed 341 malicious skills through the community skills ecosystem, targeting cryptocurrency wallets with Atomic Stealer malware.
Set against the star velocity above, those two facts describe the same phenomenon: a category growing faster than its operational practice. 386,000 stars in nine months is a lot of people installing something privileged for the first time.
Where to run one
For anyone moving from reading statistics to running an agent:
Hostinger is the practical answer — 4GB at $6.49 promotional and $11.99 at renewal, 8GB at $8.79 and $14.99, with one-click templates that remove the install step where most first attempts fail.
Hetzner has the best hardware per euro when its cost-optimized tier is in stock — check, because it has been showing as unavailable and the CPX line is €19.99 for 4GB after its June 2026 repricing.
DigitalOcean is the documentation choice at $24 for 4GB, roughly double Hostinger’s renewal rate. Its tracked link isn’t working on our end, so we’ve named it without linking.
RunPod is the one to use if you’re running a model locally rather than calling an API — hourly GPU rates from $0.34, though the arithmetic on local inference is unkind.
How we checked this
The star-velocity table is our own measurement and it’s reproducible. We queried api.github.com/repos/{owner}/{repo} on 19 August 2026 for each project, taking stargazers_count and created_at directly from the API response. Age and stars-per-month are our arithmetic: days since creation divided by 30.44, then stars divided by months. Anyone can re-run it and should, because these figures move weekly.
Its limitations, plainly:
- Stars measure attention, not adoption. No installation, deployment or retention is implied.
- Velocity since creation flatters young projects, because early growth is front-loaded by launch attention. A project nine months old with high velocity may not sustain it — that’s a reason to read the table as a snapshot rather than a trend.
- We did not obtain Docker pull counts, which this article’s brief also asked for. Docker Hub’s public API exposes pull counts inconsistently across image types and namespaces, and a partial set would have been more misleading than none.
The Gartner figures are quoted from Gartner’s own newsroom release of 26 August 2025 — that 40% of enterprise applications will feature task-specific AI agents by 2026, up from less than 5% in 2025 — together with its published best-case projection of ~30% of enterprise application software revenue by 2035 and its warning that over 40% of agentic AI projects are at risk of cancellation by 2027. We have read the release headline and reporting of it rather than the full research note, which sits behind Gartner’s paywall.
Two figures from this article’s brief that we could not verify and have therefore omitted: that roughly 31% of enterprises have at least one agent in production, and a sector spread running from 47% in banking and insurance to 14% in government. We could not trace either to a primary source, and an unsourced percentage in a statistics article is worse than a gap.
Market-size figures circulating for 2026 ($10.9–12.06bn, 44–46% CAGR) come from secondary aggregations of multiple analyst houses rather than a single primary source, so we’ve left them out of the body.
The cost figures are drawn from our own provider pricing research across this series, and the CVE and ClawHavoc figures from security reporting of the February 2026 disclosure, cited consistently across our security coverage.
What we did not do: we ran no survey, hold no proprietary telemetry, and have not installed or operated most of the projects in the table. The only original data here is the GitHub measurement, and we’ve shown exactly how to reproduce it.
The hosting links above are affiliate links.
FAQ
How fast is AI agent adoption growing?
Gartner forecasts 40% of enterprise applications featuring task-specific agents by 2026, up from under 5% in 2025 — while also warning that over 40% of agentic AI projects risk cancellation by 2027.
Which self-hosted AI agent is most popular?
By stars, OpenClaw at 386,511 as of 19 August 2026. By velocity it’s also first at roughly 43,900 stars a month — about 19 times n8n’s rate, despite not existing before late November 2025.
Is n8n still the leading automation tool?
By total stars it’s third among the projects we measured and by velocity sixth. It’s been compounding since 2019, which flatters its total and understates how much faster newer agent projects are growing.
What does it cost to run a self-hosted AI agent?
About $14–27 a month all-in: $6.49–14.99 for a 4–8GB VPS plus $2–15 of model API usage. The API is the larger and far more variable half.
Do GitHub stars mean people are actually using these tools?
No. A star is a bookmark that costs nothing and implies no installation. What the numbers reliably indicate is where developer attention is concentrated, which affects ecosystem and available help.
How many self-hosted agents are exposed on the internet?
At the CVE-2026-25253 disclosure, over 40,000 OpenClaw instances were found exposed with roughly 63% assessed as vulnerable to a one-click remote code execution flaw rated CVSS 8.8.
How much RAM does a self-hosted agent need?
4GB as the floor and 8GB once browser automation is enabled, because a headless Chromium instance wants 1–2GB by itself. That requirement is near-identical across the major projects.
