Answer Hub
The Giant Swarm Agent Platform: Everything You Need to Know
Everything you need to know about running AI agents in production — from why it matters to how it works on your own Kubernetes infrastructure, in one place.
Chapter I
Introduction
AI agents are already running inside most companies, whether IT approved them or not. The Giant Swarm Agent Platform is a way to bring that activity into the open: an open-source, self-hosted platform for running AI agents as governed, audited workloads on your own Kubernetes infrastructure, instead of as ungoverned scripts nobody can account for.
Agent Platform Key Takeaways
An AI agent platform is the infrastructure layer that lets agents act safely and continuously — not just assist someone typing faster. It runs each agent in its own isolated environment, scopes what it can access, and logs every decision, so agentic work can move from an unofficial experiment to an accountable part of production. Done well, it also gets cheaper over time as repeated agent work turns into deterministic automation.
Chapter II
Who is this for?
This page is for the people who own the decision about whether and how AI agents get to touch production systems. It's especially relevant if you are a:
- Chief Technology Officer (CTO) deciding how agentic AI fits into the company's architecture and risk posture
- Chief Information Officer (CIO) or IT lead accountable for agent sprawl, oversight and the return on AI spend
- Chief Operating Officer (COO) who needs agent-driven efficiency without new operational or compliance exposure
- Head of Infrastructure responsible for what runs on the company's Kubernetes clusters and who can access them
- Platform Lead building the internal platform that development and operations teams rely on
- Security, Compliance or Data Protection Lead who needs an answer for what an agent did, on whose behalf, and why
If your job involves deciding how AI gets access to your systems, or answering for it afterward, you're in the right place.
Chapter III
Why it matters
Most organizations aren't deciding whether to adopt AI agents anymore. They're finding out how far agents have already spread without anyone formally signing off. In Germany, 11% of companies already use AI agents and another 29% plan to, according to Bitkom's September 2026 survey of 603 companies. Employees are already using AI tools their employer never approved far more than the tools it did supply, and the gap is only widening as agentic tools get easier to wire into everyday work.
That gap has a price tag. IBM's 2025 Cost of a Data Breach Report found that 20% of organizations suffered a breach involving shadow AI, and those breaches cost $670,000 more on average than ones without it. At the same time, the upside of doing this properly is just as real: Giant Swarm has run its own production agent fleet this way since mid-2025, and turning repeated agent work into deterministic automation cut cost per run by 2.8x and tool calls by 17x. That's not a hypothetical, it's what happens once routine work stops needing a model to re-derive the same answer every time.
The uncomfortable part for most leadership teams is that somewhere in the stack, an agent is probably already running that nobody formally authorized, connected through whatever credentials an engineer had on hand. That's not an argument against agents. It's an argument for building the governance layer before an incident forces the conversation. As the platform's own origin story puts it, the real decision isn't whether your organization uses agents — it's whether that gets decided deliberately, or one unsupervised script at a time.
There's a second, quieter reason this matters: cost. Agents that just do more work, indefinitely, don't bend your AI cost curve — they add to it. The organizations that actually reduce operating costs with AI are the ones systematically turning repeated agent work into cheap, deterministic automation, rather than paying for a model to re-derive the same answer every time.
Chapter IV
How Giant Swarm approaches the Agent Platform
Giant Swarm has run production Kubernetes for enterprises across retail, telecommunications, financial services, and manufacturing for more than a decade. The Agent Platform applies that same operational discipline to agents, rather than treating them as a separate, less accountable category of software. Our approach rests on a few core ideas:
- Sovereign by default. Agents run inside your own Kubernetes environment — cloud, on-prem, edge, or fully air-gapped — with no proprietary core and no SaaS layer quietly holding your data. The principle is simple: the agent itself should run where your data already lives, even when it talks to an external model.
- Governed as infrastructure, not bolted on after the fact. Every agent gets scoped tool access, identity-based permissions, network policy, and a full audit trail from day one, the same standards we've always applied to production Kubernetes workloads.
- Built to get cheaper, not just faster. The platform's Workflow Engine watches for repeated agent behavior and turns it into deterministic, reusable workflows. That reflects a broader philosophy we hold about agentic systems in general: the goal isn't to run more AI indefinitely, it's to reach the point where routine work no longer needs a model at all, because the best AI agent is the one that makes itself redundant.
- Curated, not do-it-yourself. The agent tooling landscape changes weekly. We track it, evaluate it, and integrate what's actually worth using, so platform teams don't have to re-run that evaluation themselves every quarter.
- Run by you, or run for you. Your platform team can operate the Agent Platform with Giant Swarm's releases, maintenance and support behind it, or Giant Swarm runs it for you in your own environment, around the clock, with someone accountable for operations.
- Open, and built for how software gets made now. As agentic development becomes the default way serious engineering teams ship code, the platform and the way we build it reflect the same shift toward agent-ready, reviewable codebases that's reshaping open source more broadly.
Chapter V
FAQs
What is the Giant Swarm Agent Platform?
Think of it as the difference between letting engineers duct-tape AI into your systems one integration at a time, versus giving them a proper piece of infrastructure to build on. The Agent Platform is an open-source, self-hosted platform for running AI agents as governed Kubernetes workloads — every agent isolated, every tool call scoped and audited, on infrastructure you already control. It has three parts. Agent Runtime is where domain experts build and share their own agents from building blocks IT has approved, and reuse what other teams built. Identity and Access makes sure an agent acts only with the permissions of the person or role behind it, with every action traced to the agent and the person. Model Runtime puts frontier providers, European AI clouds or your own open-weight models behind one gateway, with every call attributed to a person and a team. Giant Swarm built this the hard way, by running its own fleet of production agents on this same setup since mid-2025 before packaging it for customers, so this isn't a whiteboard concept — it's infrastructure that's already been through the kind of 3 a.m. incidents that expose what doesn't work.
How do you run AI agents on Kubernetes in production?
The short answer: you treat an agent like any other production workload, not like a script someone's laptop happens to be running. On the Agent Platform, every agent is deployed as a declarative Kubernetes resource — reviewed, versioned, and rolled out through GitOps, the same way your other production services are. That means your platform team's existing muscle memory (pull requests, rollback, policy enforcement) applies to agents from the first day they run, rather than agents living in some ungoverned side-channel. Kubernetes is already where AI workloads are going: 66% of organizations hosting generative AI models use Kubernetes to manage some or all of their inference workloads, according to the CNCF's 2025 Annual Cloud Native Survey. The agent itself, not just the model it talks to, runs inside your cluster — which matters more than it sounds, because the part that touches your logs, your infrastructure, and your customer data is exactly the part you want under your own operational control, regardless of which LLM provider sits behind it.
How can we give AI agents access to internal tools without creating new security risks?
This is usually where teams get nervous, and for good reason — most organizations today connect agents to internal tools one integration at a time, each with its own credentials and its own blind spot. That approach doesn't scale, and it's exactly the setup that leaves security teams with no good answer when someone asks who authorized an agent to do what. The Agent Platform's answer is a single governed gateway sitting between every agent and every tool: single sign-on through your identity provider, per-user scoped access, allowlists, and a complete audit trail. Critically, the gateway doesn't hold shared service credentials at all — agents act on behalf of the signed-in user via OAuth on-behalf-of tokens, so there's no standing “agent account” with broad access sitting around waiting to be misused. This is the same problem we described when agents started connecting to everything across the industry with nobody governing the connections — the platform exists specifically to close that gap before it becomes an incident.
Does the Agent Platform work with our existing infrastructure?
Yes, and this is by design rather than by coincidence. The platform runs on any CNCF-conformant Kubernetes cluster, version 1.33 or newer — cloud, on-prem, edge, or air-gapped, whether Giant Swarm manages it or your own team does. You bring your own models — from a model provider, a European AI cloud, or open-weight models you host yourself, for example with vLLM or Ollama — your existing MCP servers plug straight into the platform's governed gateway, and the UI can sit inside an existing Backstage portal if that's already how your teams navigate internal tooling. If you don't have Kubernetes yet, that's not a blocker either — a single server is enough to get started. Nothing about adopting the platform requires ripping out what you've already built; it's meant to sit on top of it.
Do we get audit trails and scoped access control?
Yes — and on this platform, that isn't an add-on feature, it's the operating model. Every tool call an agent makes is traced back to the person it was acting on behalf of, so “who authorized this and what did it touch” has an actual answer instead of a shrug. Access is scoped per user: each person, and each agent acting for them, only ever sees the tools they're individually entitled to use. This matters because the alternative — a shared service account with broad standing access that any agent can invoke — is precisely the pattern that turns a convenient shortcut into a breach headline. Given how much shadow AI already costs organizations when it goes wrong, building audit and scoping in from the start isn't a nice-to-have, it's what makes running agents in production defensible to your own security team.
Can we run AI agents in regulated or air-gapped environments?
Yes. The platform is fully self-hosted, which means it runs on the factory floor, at the edge, and in fully air-gapped environments with no outside connection at all — not just in a conventional cloud account. Your data, your models, and every agent interaction stay inside your own perimeter throughout. That's what actually makes agents viable for regulated industries and for production systems where a call home to a third-party service isn't an option in the first place. Regulations such as NIS2, DORA and the EU AI Act all raise the bar on showing who did what in your systems, and a per-user audit trail of every agent action is the kind of evidence they ask for. It follows the same logic behind why the agent itself needs to run where your data already lives, rather than shipping your logs and internal context somewhere else just so a model can look at them.
The platform is open source. What do we pay for?
Fair question, and the honest answer is: not the software itself. The full platform is open source with no feature gates hidden behind a paywall — anyone can go pull the open-source install today and it will start empty. What you're paying for, if you become a customer, is the head start and the ongoing curation: a platform pre-filled with curated agents, skills, and tool integrations that Giant Swarm has already refined on its own agent fleet, delivered as tested and signed images, plus support and optional full management. That head start matters because few teams have spare engineers to build it themselves: 78% of German companies report a shortage of IT specialists, according to Bitkom's September 2026 labour market survey. It's a similar economic shift to what's happening across agentic software more broadly — as teams increasingly build and maintain software with agents doing the implementation work, the value shifts from “who wrote the code” to “who's curating, reviewing, and keeping the whole thing trustworthy over time.” That curation is the paid part.
How is the Agent Platform different from hyperscaler agent offerings or building our own platform?
Every option here has a real trade-off, so it's worth being blunt about them. Hyperscaler and SaaS agent offerings get you moving fastest, but your agents end up depending on one vendor's cloud, jurisdiction, and pricing, and your data leaves your perimeter to get there. Building your own platform keeps you in control, but it means someone on your team has to continuously evaluate and integrate an agent tooling landscape that changes weekly, indefinitely, with no end date. The Agent Platform runs on that same open landscape, kagent, agentgateway, and muster, not a proprietary layer, so you're never locked into our roadmap either. It is meant to be the middle path: production-ready now, fully open source and running inside your own environment, with any model or tool swappable whenever you want. It also reflects a philosophy we think most vendors won't tell you: the goal isn't to maximize how many agents you're running, it's to reach the point where routine agent work gets converted into cheap, deterministic automation instead of an ever-growing token bill — which is much harder to do if you're locked into a single vendor's metering.
Does the Agent Platform meet digital sovereignty and data-residency requirements?
Yes, and this is arguably the core design constraint the whole platform is built around, not a checkbox added afterward. Everything is open source and self-hosted on your own infrastructure. You can work with models you host yourself, so your data never has to leave your perimeter and no third-party cloud is a required dependency. It runs air-gapped where that's a hard requirement, and it comes from a European vendor with ISO 27001 certification and over a decade of production Kubernetes experience behind it, which matters if data residency and jurisdiction are part of your compliance picture. It also speaks to the biggest worry companies have about AI: in Germany, 66% of companies that use AI name data protection requirements as an obstacle, ahead of legal uncertainty at 56% (Bitkom, September 2026). The underlying principle is the same one behind Giant Swarm's broader approach to agent infrastructure: the agent needs to operate where your data already sits, under your governance, not the other way around.
Closing thoughts
AI agents are already inside your organization, whether or not anyone signed off on it. The question the Agent Platform is built to answer isn't whether to allow that — it's how to make it safe, auditable, and cost-effective once it's already happening. Giant Swarm brings the same production-Kubernetes discipline it has applied for over a decade to this newer layer of infrastructure.
Talk to us about the Agent Platform or explore the Agent Platform in more detail.


