Your engineers are already running AI agents. Not as an approved rollout, but as something that crept in over the past year, wired into whatever tools got a task done.
It's already widespread: IDC found that 56% of employees use AI tools their organization never provided, versus only 23% who use what their company supplied (IDC Global Employee Experience Survey, 2025). It's already expensive: IBM's 2025 breach report found that one in five organizations were breached through shadow AI, and that breaches involving it cost $670K more than those with little or none (IBM, 2025). And it's already changing what reaches production: Faros AI's study of 22,000 developers found 243% more incidents per pull request as AI adoption deepens (Faros AI, 2026).
That is not an argument for banning agents. Amazon's Q Developer agents migrated 30,000 applications from Java 8 to 17, saving an estimated $260M a year and 4,500 developer-years of work (AWS, 2025). EY now runs 150 agents supporting 80,000 tax professionals on compliance work (EY, 2025). The upside is real. Your organization already uses agents. What you still get to decide is how, before your engineers decide it for you, one unsupervised script at a time.
Tools make people faster. A platform lets agents act.
Most of what gets called "AI agent tooling" today (workflow builders, IDE copilots, chat wrappers, the coding agents you run from your terminal) is built to make one person faster at their keyboard. That's genuine value, and it stays exactly there: with the person, at their machine, triggered by their hand. None of it was built to run unattended, hold credentials safely, or leave a trail anyone can check.
An agent platform is a different layer of infrastructure. Agents act on events rather than clicks: an alert, a ticket, a scheduled job. Each one runs isolated, with scoped access to only the tools its task needs, and every decision it makes is logged. That's what separates AI that helps someone type faster from AI you can trust to run unattended in production, with an audit trail for everything it did.
We didn't design this on a whiteboard
Giant Swarm has run production Kubernetes for enterprises for more than a decade, for companies like adidas, Vodafone, Deutsche Telekom, Börse Stuttgart, and EGGER, across retail, telecommunications, financial services, and manufacturing. Since summer 2025, we've been running our own fleet of agents on the same infrastructure, triaging incidents, reviewing code, and clearing backlog that nobody had time to handle by hand. The Agent Platform is that infrastructure, packaged so you can run it too.
It's sovereign by design. Everything runs inside your own environment, on any Kubernetes cluster you already operate, whether that's cloud, on-prem, edge, or fully air-gapped. There's no proprietary core, and no SaaS calling home with your data. If it runs Kubernetes, it runs the platform, under the same controls and standards we've always applied to your clusters.
It also gets cheaper and more reliable the longer you run it. Every time an agent completes a task, the platform's Workflow Engine captures how it got there and automatically turns the patterns that repeat into deterministic MCP Workflows. The next run leans on that codified path instead of the model, so it needs fewer tokens and less time. In our own operation, that loop already shows up as 2.8x lower cost per agent run and 17x fewer tool calls, with an average of around 500 agents running in parallel across the company. The numbers keep improving because the platform learns from every run.
What it looks like once it's running
The shape changes with the job, but the pattern holds. In software engineering, it's the difference between one engineer's AI-assisted laptop and a whole team's output: our own engineering org shipped 318% more pull requests per month this spring than a year earlier, with the same headcount. In IT operations, it's an agent that triages an incident and proposes a documented fix within 2 minutes, instead of the hours it takes to dig through logs and dashboards by hand, with a human still deciding what ships. In compliance, it's continuous evidence collection instead of a frantic scramble before an audit deadline.
This doesn't replace judgment. It removes the parts of the job that were never about judgment to begin with.
Three commitments behind every decision
Sovereign. Open source, with the freedom to swap any model, harness, or tool whenever you want. Your data never leaves your control, and there's no lock-in to work around later.
Curated. The agent landscape changes weekly. We track it, evaluate it, and integrate what's worth using, so your team doesn't have to.
Enterprise-ready. Multi-tenant, secure, and auditable from the first agent you run, rather than bolted on after something goes wrong.
This isn't new territory for us. It's the same bar we've held ourselves to on Kubernetes since 2014, applied to a newer layer of infrastructure.
Become a design partner
We're opening the Agent Platform to a small group of design partners ahead of the broader launch later this year. These are teams already running agents, or about to, who are hitting the same wall we did: orchestration, governance, and cost, all at once.
Early access comes on founding-customer terms. The platform gets deployed on your own infrastructure, with our engineers working alongside yours. The platform is still in active development, and that's deliberate: as a design partner, the gaps are where you get direct influence over what we build next.
Your engineers are already deciding how agents run in your environment. The open question is whether you're making that decision with them.
Start the conversation at giantswarm.io/agent-platform
