What it actually costs to run 1,000+ AI agents on your own infrastructure

This is not a license, it is a team

Most build-versus-buy conversations about AI agents start in the wrong place: the price of the software. Open-source runtimes like OpenClaw and Hermes are free, so it looks cheap. But the runtime is the smallest part of the bill.

What actually costs money is everything around the agent: keeping a fleet of them alive, recovering them when they crash, packing many safely on shared hardware, versioning their state, wiring them into your tools, and being on-call when something breaks at 3am. That is not a purchase, it is an operating model. And once you write it down as headcount and months, the honest number surprises most teams.

We run this system for a living, so this is not a sales scare. It is the breakdown we wish someone had handed us.

What those engineers actually build

Before a single agent is production-ready at scale, someone on your team has to build and own each of these. None of them is optional once you are past a handful of agents and expected to stay online.

• Self-healing: detect a crash in seconds, restart in-pod, recreate the pod if that fails, restore known-good state if that fails, and alert a human if all else fails. Getting this to catch crashes in under 60s and restore in under 90s is months of work.
• Safe density: a memory guard that packs many agents on a node and kills by priority before the node runs out of memory. Without it, one spike takes down every agent on the box.
• Versioned state: a filesystem where every file is versioned with point-in-time restore and rollback, so a bad run or a delete is recoverable and hot-reloads the running agent.
• Integration layer: 1,000+ app connections, OAuth flows, per-instance scoped credentials encrypted at rest, browser automation with captcha handling and proxies, plus email and voice per agent.
• Upgrade pipeline: OpenClaw and Hermes ship new versions constantly. You need to build, test and roll them out per agent without breaking anyone.
• Observability and governance: fleet-wide monitoring, audit, kill-switch and priority controls, so you can actually run 1,000 agents instead of guessing.

The team math, done honestly

The strict minimum to stand this up is about two DevOps or platform engineers plus one full-stack engineer for the control plane and dashboard. That gets you something that runs.

But "runs" is not "stays online." A fleet of 1,000+ agents that your business depends on needs a 24/7 on-call rotation, and you cannot run a sustainable rotation on two people: no holiday cover, no sick cover, burnout in a quarter. A real rotation is four to five engineers. Add security and compliance review for a regulated deployment and it climbs again.

In loaded cost, an infrastructure engineer runs roughly 120k to 150k euros a year in Europe. So the honest figure lands around 500k to 800k euros a year in dedicated headcount, plus six to twelve months before your first fleet is production-hardened and compliant. That is the number that belongs in the budget meeting, not the zero on the open-source license.

The part nobody budgets: the treadmill

Here is the cost that quietly doubles the bill: it never ends.

An agent platform is not a project you finish. OpenClaw and Hermes release new versions on a fast cadence, and each one can change behavior your setup depends on. The 1,000+ integrations rot: APIs change, OAuth scopes shift, a vendor deprecates an endpoint, and something silently breaks. The density and self-healing logic needs continuous tuning as your workloads change.

So the team you staffed is not a one-time build cost, it is a permanent maintenance tax on a system that does not differentiate your business. A managed platform amortizes that treadmill across every customer. In-house, you pay for all of it, alone, forever.

If you already have a platform team

Large enterprises read the headcount and shrug: they have hundreds of engineers. Fair. For you the argument is not "you cannot staff it," it is opportunity cost.

Your platform team's time is your scarcest asset, and rebuilding agent infrastructure that already exists and evolves every week is the definition of undifferentiated heavy lifting. Every sprint spent on self-healing and integration plumbing is a sprint not spent on what actually differentiates your business. The question is not whether you can build it. It is whether reinventing this particular wheel is the best use of the best engineers you have.

The alternative: buy the socle, keep it on your metal

This is the part where the honest math points somewhere. Molted is the exact stack described above, self-healing, safe density, versioned state, 1,000+ integrations, upgrade pipeline, governance, built once and maintained for you.

And it does not force you into someone else's cloud. Molted runs on managed clusters, or fully on-premise on your own infrastructure, or air-gapped, so your data never leaves and you still skip the platform team. You get the fleet without the on-call rotation, and your engineers stay on the work that differentiates you.

That is the whole trade: build a permanent internal product with a permanent team, or run the same thing on your hardware and pay for outcomes instead of headcount.