Architecture Overview
Autoheal runs as three tiers:
- always-on platform services that hold your data and orchestrate agents
- long-running runners that host execution
- ephemeral sandboxes where tool commands actually run
The governance control plane described in the Overview, covering agent definitions, budgets, identity and approval policy, runs on the platform services tier.
The same three tiers exist whether you run Autoheal as SaaS or in your own cloud (BYOC). Only where they run changes. This page describes what each tier is for and what it stores, then the two deployment models.
The three tiers
The diagram below shows the three tiers as they run today in Autoheal's SaaS — on Amazon EKS in Autoheal's AWS account. Connected BYOC runs the same architecture in your own account, with private instead of public ingress. Airgapped BYOC runs the same tiers on Docker Compose on hosts you provision (see Deployments).
Users reach the always-on platform services over SSO through a public, SSO-gated load balancer. The platform services tier (on EKS) holds all durable state and dispatches work to a long-running runner. The runner hosts ephemeral sandboxes. The sandboxes run tool commands against your systems, within each agent's tool filters, and write results back. Inference is routed to the model provider best suited to each task — some in-account, some direct. The application, runner, and all your data sit inside Autoheal's AWS account, isolated per tenant.
The runner is a deliberate boundary — the code that executes commands against your environment never sits alongside the services that hold your durable state.
| Tier | Lifecycle | Runs | Stores |
|---|---|---|---|
| Platform services | Always on | Web UI, API, agent orchestration, identity, workflow | Everything durable — investigations and run traces, agent specs, memories and skills, the Engineering Context Graph, users and roles, audit logs, integration config. PostgreSQL is the system of record. Object storage holds artifacts. Redis is a transient cache/stream. Secrets are in a dedicated store. |
| Runner | Long-running, autoscaled | Sandbox lifecycle, scoped-credential brokering | Transient only — cached repos and working files, reaped automatically. No system of record, no database. |
| Sandbox | Per run, inside a runner | Tool commands: CLIs, kubectl, MCP tools | Nothing durable. The filesystem is destroyed at cleanup, and outputs are written back to platform services. |
Terminology. An investigation is the durable unit of diagnostic work for one issue: an alert, an incident, or a question you pose. An investigation may execute over one or more runs (the first run plus any later resume or follow-up). Each run gets one sandbox, created when the run starts and destroyed when it ends. So an investigation uses one sandbox at a time and a fresh one on each resume, never many at once. One runner hosts many sandboxes over its life.
Data flows
Where the platform runs differs by model. In SaaS, the platform services tier runs in Autoheal's account, so your data flows out to it (SSO-gated, isolated per tenant). In BYOC, it runs in your account, so data leaves only on the egress paths you choose. What's constant is how the agent touches your data and what paths exist into the platform:
| Traffic | Direction | Destination / source | Carries |
|---|---|---|---|
| LLM inference | Outbound | The model configured for your deployment — an Autoheal-supported provider, in-account (Bedrock / Azure OpenAI / AI Foundry) or direct (OpenAI / Anthropic) | Prompts and context: investigation and alert context, tool output, and excerpts of the data the agent reads |
| Integration calls | Outbound | Your observability, code, cloud, and database endpoints | Read-only queries by default. Writes occur only where the integration has write tools and access is granted, gated by any matching approval policy |
| Alert ingestion | Inbound | Your monitoring (Alertmanager, Dynatrace, …) posts to an authenticated webhook | Alert payloads: labels, annotations, severity |
| User access | Inbound | Your users' browsers, SSO-gated | Sign-in and app requests |
The agent's access to your systems is outbound and read-only by default. Many integrations ship only read tools. Where write tools exist, access is granted explicitly, and approval policies can pause each write for a person to approve. Every inbound path is authenticated and configured by you. In BYOC, all of this stays on private networking with no public inbound. With an in-account model and in-network integrations, nothing leaves your account at all (see What leaves your environment).
Durability and failure handling
Both the investigation and the sandbox lifecycle run as durable Temporal workflows whose state lives on the platform services tier, not on any single runner. A runner is disposable. If one is terminated mid-run, the workflow re-places the sandbox on a surviving runner and rebuilds it from spec (repositories re-cloned, init re-run). A command that was in flight when the host was lost is not silently retried. Its outcome is unknown, so it is marked indeterminate and the agent decides whether to re-run it. Re-placement is bounded (a few attempts). Past that, the run is marked failed rather than retried forever. Nothing durable lives on the runner, so losing one costs a rebuild, not data.
Deployments
Autoheal offers two deployment models: SaaS, or BYOC — the same three tiers, either in Autoheal's account or in your own.
Autoheal works with your team on account, network, and compliance specifics before the first install.
| SaaS | BYOC: connected | BYOC: airgapped | |
|---|---|---|---|
| Runs in | Autoheal's cloud account | Your cloud account (EKS or AKS) | Hosts you provision, on an isolated network (Docker Compose) |
| Data at rest | Autoheal's account, isolated per tenant | Your account and region | Your hosts |
| LLM inference | Autoheal-managed provider accounts | The provider you choose (in-account or direct) | A model endpoint reachable from your isolated network |
| Upgrades | Automatic | You control the window | Your team imports each release bundle |
| Best for | Fastest start, most teams | Data residency or contractual requirements that keep data in your account | Environments with no path to the internet or to Autoheal |
SaaS
- Runs in Autoheal's own infrastructure. You access it in your browser over SSO.
- Connect your tools as integrations, then start building agents and running investigations.
- Autoheal hosts, operates, and upgrades everything — there is nothing for you to run.
- Your data is logically isolated per tenant.
- Most teams should start here.
This is exactly the deployment shown in the diagram above: the three tiers running on Amazon EKS in Autoheal's AWS account. Ingress is public and SSO-gated, and outbound access to your environment is read-only by default.
BYOC
For customers with strict security and data-protection requirements, Autoheal offers BYOC in two modes, connected and airgapped. Connected BYOC:
- Runs on Amazon EKS or Azure AKS, inside your own Kubernetes cluster — a single-tenant, shared-nothing installation.
- The application, orchestration, identity, and all data stay entirely within your cloud account.
- Connections to your data sources flow over private, encrypted network paths.
- There is no public inbound — no Autoheal-hosted service reaches into your environment at runtime.
Connected BYOC runs the same three tiers, the same skeleton as the SaaS diagram above, but the whole box is your account. Ingress is private with no public inbound. Data leaves only on the egress paths you choose, such as a direct model provider or an external SaaS integration. You can run every install and upgrade yourself. If you'd rather Autoheal operate it, you grant a scoped, revocable role that Autoheal assumes only for the install window. Either way, Autoheal keeps no standing access. See BYOC for the full AWS (EKS) and Azure (AKS) topology.
Airgapped BYOC runs the same tiers on Docker Compose on hosts you provision. It runs from an imported image bundle, with no path to Autoheal and no Autoheal operator access. See Connected and airgapped.
Who does what: your platform team provisions and manages the infrastructure (cluster or hosts, network, managed data, IAM) and runs installs and upgrades. In connected BYOC, you can instead grant Autoheal a scoped, revocable role to run the install or an upgrade. In airgapped BYOC, your team imports and runs each release bundle.
Get started
Both models start with a conversation. Contact us at autoheal.ai or support@autoheal.ai.
- SaaS — we configure your tenant and connect your tools. Your team runs investigations in the browser.
- BYOC — onboarding covers the mode (connected or airgapped), sizing for your load, the network and egress allowlist for connected installs, and your LLM and identity configuration. It also covers the install, which we run together with your platform team.