Tune Agent Substrate

Learn how to size a WorkerPool, configure snapshot storage, and keep Workers on the sandbox class that kagent requires.

kagent runs every agent on Agent Substrate. A default installation creates one WorkerPoolWorkerPoolA Kubernetes custom resource declaring how many Workers to keep running and which sandbox class they use. An operator must provision one before any Agent can compile.Learn more named kagent-default, holding a single Worker on the gvisor sandbox class, and it writes snapshots to the object store that the Agent Substrate chart bundles.

This single Worker reports capacity for 1000 Actors, so a default pool serves many concurrent SessionsSessionA running conversation with one Agent. Unlike the resources it is built from, a Session is not a Kubernetes resource: kagent's gRPC API creates it and its PostgreSQL database tracks it.Learn more before it runs out of slots. Three settings still need tuning before the installation takes real traffic: how many Workers to run, where snapshots are stored durably, and what resource requests the Worker pods set. This page covers each setting, and the steps for moving a pool to the microvm sandbox class.

Before you begin

Install kagent, and confirm that your installation sets controller.grpc.reflection=true. Reflection lets a gRPC client discover the controller’s methods without a local copy of kagent’s protocol buffer definitions.
Confirm that you have administrative access to the cluster, because this guide scales WorkerPools and reads cluster-scoped resources.
Install grpcurl. GetSubstrateStatus has no kubectl or kagent command equivalent, so calling it directly is the only way to see the full runtime state.

Inspect the runtime

Read the current state of the runtime before you change it. The rest of this guide uses three values from this section: the pool name and its READY count, the snapshot path that Actors are writing to, and the Worker image that the pool runs.

Check the WorkerPool’s Worker counts. These counts show whether the pool is producing the Workers that you asked for.
```
kubectl get workerpools -n kagent
```
Example output:
```
NAME             DESIRED   REPLICAS   READY   AGE
kagent-default   8         8          8       3h
```
DESIRED is the replica count that you asked for, REPLICAS counts the Worker pods that exist, and READY counts the pods that are serving. A gap between DESIRED and REPLICAS means the controller is not creating pods. A gap between REPLICAS and READY means the pods exist but are not starting. Raising the replica count helps only when all three counts agree.
Port-forward the controller, and leave the command running in a second terminal.
```
kubectl port-forward -n kagent svc/kagent-controller 8083:8083
```

Call GetSubstrateStatus, which joins the WorkerPools, ActorTemplatesActorTemplateThe compiled, immutable definition that the kagent controller produces from an Agent. Every Actor is created from one.Learn more, ActorsActorThe sandboxed unit of compute, provided by Agent Substrate, that runs a Session's conversation loop. Every Session is backed by one.Learn more, and Workers that kagent knows about into one response.

grpcurl -plaintext -d '{"namespace":"kagent"}' \
  localhost:8083 kagent.api.v1alpha1.SystemService/GetSubstrateStatus

Example output, abbreviated to one Worker and one Actor:

{
  "enabled": true,
  "workerPools": [
    {
      "namespace": "kagent",
      "name": "kagent-default",
      "replicas": 8,
      "ateomImage": "ghcr.io/kagent-dev/substrate/ateom-gvisor:v0.4.0-alpha1"
    }
  ],
  "actors": [
    {
      "actorId": "792f8f90-72a9-49e8-b720-cb3b96fd9a8b",
      "atespace": "ate-golden",
      "status": "Suspended",
      "actorTemplateName": "assistant-kagent-2a786e2db23a",
      "latestSnapshot": "s3://ate-snapshots/kagent/atespaces/ate-golden/actors/2d921676-80ad-48d3-8c2a-4059ef74da33/snapshots/503a07e3-ec49-433c-abdf-b2bb3cb36d5d"
    }
  ],
  "workers": [
    {
      "workerNamespace": "kagent",
      "workerPool": "kagent-default",
      "workerPod": "kagent-default-787547df77-4kvqf",
      "ip": "10.244.0.104"
    }
  ]
}

A Worker entry gains an actorId only while an Actor occupies it, so the Workers list also shows which capacity is busy.

Size a WorkerPool

The replica count is the only capacity dial on a pool, and setting it well depends on knowing what actually consumes a Worker.

A WorkerWorkerA pre-started, sandboxed pod that hosts at most one Actor at a time. Actors are multiplexed across a pool of Workers over time.Learn more hosts many Actors at the same time, and it holds each one only while a turn is running. kagent suspends an Actor at every turn boundary and frees the Actor’s share of its Worker, as described in Suspend and resume. An idle SessionSessionA running conversation with one Agent. Unlike the resources it is built from, a Session is not a Kubernetes resource: kagent's gRPC API creates it and its PostgreSQL database tracks it.Learn more therefore occupies no Worker capacity at all.

Each Worker reports an Actor capacity when it starts, fixed by the Worker build rather than by a field on the WorkerPool. The gVisor and micro-VM Workers that Agent Substrate ships both report a capacity of 1000 Actors, and no setting raises or lowers it.

Agent Substrate checks this capacity on every placement. It weighs a Worker’s spare CPU and memory only when the ActorTemplate states how much CPU and memory one Actor needs. An ActorTemplate that the kagent controller compiles states no per-Actor CPU or memory, so a pool places a Session’s Actor on a Worker’s available slots alone, regardless of how loaded the Worker’s node already is. Worker resource limits still matter, because they govern how Kubernetes packs Workers onto nodes.

Size the pool for the number of turns that run at the same time, not for the number of Agents or Sessions that you have created. A cluster with hundreds of Sessions that are each used occasionally needs far fewer Workers than the session count suggests.

Scale a pool in place with the standard Kubernetes scale command, because the WorkerPool exposes a scale subresource. Use the pool name that you read in Inspect the runtime.

kubectl scale workerpool kagent-default -n kagent --replicas=8

To change these settings in your Helm values instead, run an upgrade of your kagent installation with the following settings:

substrateWorkerPool:
  replicas: 8

Recognize an undersized pool

Agent Substrate reports pool capacity through its metrics rather than through an error that reaches the caller. Each Worker publishes a state of idle, partial, at_capacity, or unschedulable, which together show how much of a pool is in use. A placement that finds no suitable Worker records a no_capacity outcome. Agent Substrate treats this outcome as a normal result rather than a failure, so it carries no error type and raises no alert on its own. Watch the share of Workers at at_capacity and the rate of no_capacity placements to tell a pool that is merely busy from one that is too small.

None of those signals reach the caller. A turn that cannot get a Worker waits for one, and a turn that waits too long fails with a generic timeout that names neither capacity nor the pool.

ERROR:
  Code: Internal
  Message: actor "ai-01a087c0-1d72-775a-9a32-566acac7b685" request timed out

Because short turns release their capacity quickly, a pool under mild pressure absorbs the load and only sheds the requests that wait past the deadline. The result is intermittent timeouts under concurrency rather than a clean failure. Neither the kagent controller log nor the Agent Substrate API log records a capacity message when this happens, so confirm the diagnosis from the at_capacity and no_capacity signals and from the pool’s READY count. If the pool is under pressure, raise the replica count.

Set resource requests on Workers

Generated Worker pods carry no resource requests or limits unless you supply them, and the effect on scheduling is easy to miss.

A Worker with no limits reports its node’s entire allocatable capacity as its own. Every Worker on that node reports the same figures, so Agent Substrate cannot use resource capacity to place Actors, and Kubernetes cannot stop you from packing more Workers onto a node than it can carry. Set requests and limits on any pool that carries production traffic.

substrateWorkerPool:
  template:
    resources:
      requests:
        cpu: 500m
        memory: 1Gi
      limits:
        cpu: "2"
        memory: 4Gi

The same template block also accepts nodeSelector, tolerations, priorityClassName, and nodeAffinity, so you can keep Workers on nodes that you set aside for them.

Configure snapshot storage

Agent Substrate writes every suspended Actor to an object store, so the bucket stores the Session state that a resume restores. Two places decide where a snapshot lands:

In the Agent Substrate Helm values, atelet.storageBackend selects the storage client and rustfs.enabled decides whether the chart supplies credentials for it. These two values cover both the components that write snapshots and the ones that manage them.
In each HarnessHarnessA Kubernetes custom resource defining how an agent is allowed to run: its runtime, workload image, and WorkerPool and snapshot storage. An Agent pairs it with the AgentTemplate it runs.Learn more, spec.substrate.snapshotPolicy.location, names the bucket and prefix that its Actors write under. The field is required, and it must point into a bucket that the selected client can reach.

The bundled object store

Because rustfs.enabled defaults to true, the Agent Substrate chart installs an in-cluster object store and points the atelet DaemonSet and the ate-api-server Deployment at it. The store claims a 1Gi PersistentVolumeClaim, and it accepts an access key and secret key that the chart sets to published defaults.

Important

The bundled store is not durable. Its credentials are shared defaults that every installation receives. Snapshots store agent Session state, so move a production installation to your own bucket before it carries real traffic.

Move snapshots to an external bucket

Turn the bundled store off and select the backend in your Agent Substrate Helm values. Setting atelet.storageBackend to s3 selects the S3 client, and setting it to gcs selects Google Cloud Storage. Agent Substrate matches only the s3 value, so every other value, an empty one included, selects Google Cloud Storage.
```
rustfs:
  enabled: false
atelet:
  storageBackend: s3
```
Supply the credentials that your backend needs. Setting rustfs.enabled to false stops the chart from setting the AWS_* variables, so neither component receives an endpoint or a key from Helm. Each storage client reads its credentials from the pod’s own environment instead.
- For Google Cloud Storage, the client uses application default credentials. Bind the components’ ServiceAccounts through Workload Identity Federation for GKE. The chart renders no Helm value for Google credentials, so set nothing further.
- For Amazon S3, the client reads the standard AWS environment and uses a role that the pod assumes. No static key is needed.
- For an S3-compatible store, pass two variables and the static credentials through atelet.extraEnv. AWS_ENDPOINT_URL names the store’s API endpoint, which the client otherwise resolves to Amazon S3. AWS_S3_USE_PATH_STYLE puts the bucket name in the request path instead of the hostname, which most S3-compatible stores require.
```
atelet:
  extraEnv:
    - name: AWS_ENDPOINT_URL
      value: https://objects.example.com
    - name: AWS_S3_USE_PATH_STYLE
      value: "true"
```

Point each Harness at the new bucket.

spec:
  substrate:
    workerPoolRef:
      name: kagent-default
    snapshotPolicy:
      location: s3://ate-snapshots/kagent/

Run a turn on an agent, suspend its Actor, and compare the latestSnapshot path in the response against the location that you set. Because the backend and the location are configured separately, a mismatch between them breaks resuming an Actor from its snapshot later.

How the chart reaches the components

The atelet DaemonSet writes a snapshot, and the ate-api-server Deployment manages the snapshot. The chart renders atelet.storageBackend as the ATE_STORAGE_BACKEND variable on the DaemonSet and the Deployment alike, so one Helm value keeps the two in agreement.

Reading these variables off a running pod helps when a snapshot fails to write. Edit the Helm values rather than the pod environment because the next Helm upgrade overwrites a direct edit.

Variable	Purpose
`ATE_STORAGE_BACKEND`	`s3` selects the S3 client. Any other value, including unset, selects Google Cloud Storage.
`AWS_REGION`	Region for the S3 client.
`AWS_ENDPOINT_URL`	Endpoint for an S3-compatible store that is not Amazon S3.
`AWS_S3_USE_PATH_STYLE`	`true` selects path-style addressing, which most S3-compatible stores need.
`AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY`	Static credentials for the S3 client.

Snapshot layout in the bucket

Agent Substrate builds a path per Actor under the Harness prefix, so one bucket serves many Harnesses without collision. The <actor-uid> segment is the Actor’s metadata.uid, not the actorId that GetSubstrateStatus reports. Both values are UUIDs, so read latestSnapshot off the Actor rather than matching a path to an Actor by eye.

s3://ate-snapshots/kagent/atespaces/<atespace>/actors/<actor-uid>/snapshots/<snapshot-uid>

A checkpointCheckpointA durable pin on the snapshot that a Session most recently suspended to, and a record of how far its transcript had advanced. Not a new state: tagging copies the snapshot so that Agent Substrate does not collect it, and a second Session can be forked from it.Learn more uses a different path. kagent creates an Agent Substrate TagTagAn Agent Substrate resource that gives one snapshot a stable, human-meaningful name, so callers do not need to track Substrate's internal snapshot identity. It also acts as a retention pin: Agent Substrate does not collect a snapshot while a tag names it, and only the tag's visibility scope can change afterward.Learn more for each checkpoint, and tagging copies the Actor’s snapshot into the Tag’s own prefix. A Tag names exactly one snapshot, so its objects sit directly under tags/ with no snapshots/ segment. The <tag-uid> segment is the Tag’s metadata.uid, so nothing in the path records the checkpoint-<id> name that kagent gave the Tag.

s3://ate-snapshots/kagent/atespaces/<atespace>/tags/<tag-uid>

A backup or lifecycle policy must cover both prefixes, because one scoped to actors/ omits every checkpoint and understates how much the bucket stores.

Keep pools on the sandbox class they serve

kagent configures the sandbox class for you. A default installation runs the gvisor class, and every ActorTemplateActorTemplateThe compiled, immutable definition that the kagent controller produces from an Agent. Every Actor is created from one.Learn more that kagent generates inherits the class of the pool that hosts it. Edit this setting only to move a pool to the microvm class or to change a pool’s Worker build. For what each class isolates, see Sandboxing.

One rule governs the change. A pool’s class, its Worker image, and the cluster’s SandboxConfig must all name the same sandbox class, and Agent Substrate rejects the Actor rather than falling back when they disagree.

Set a pool’s sandbox class

Set the class and the matching Worker build in your kagent Helm values.

substrateWorkerPool:
  sandboxClass: gvisor
  workerImage: "ghcr.io/kagent-dev/substrate/ateom-gvisor:v0.4.0-alpha1"

Field	Description
`sandboxClass`	The sandbox runtime family for this pool’s Workers, and for every ActorTemplate that kagent generates for the pool. Omitting the field selects `gvisor`.
`workerImage`	The Worker build that provides the runtime. `ateom-gvisor` serves the `gvisor` class and `ateom-microvm` serves `microvm`. A build that does not match its pool’s class never registers a Worker.

Confirm that the cluster has a SandboxConfig for this class. kagent names the configuration directly rather than resolving a default, so a gvisor pool needs a config named gvisor-default and a microvm pool needs a config named microvm.
```
kubectl get sandboxconfigs
```
Example output:
```
NAME             CLASS    AGE
gvisor-default   gvisor   26h
microvm          microvm  26h
```
A default installation creates only the gvisor class, because the Agent Substrate chart applies sandboxconfig-gvisor.yaml and kagent installs nothing for microvm. The microvm configuration and its runtime assets are staged out of band by the Agent Substrate installer, such as through hack/install-microvm-deps.sh --install.
Verify that the pool’s Workers came up on the build that you set. The ateomImage field in the runtime response reports the image that each Worker is running, which the WorkerPool resource calls workerImage. Compare the two, and confirm in the output that the pool reports its full replica count as ready.
```
kubectl get workerpools -n kagent
```

Diagnose a class mismatch

Each of the three settings fails differently when it disagrees with the others.

A Worker whose pool class does not match an ActorTemplate’s class reports RevisionInvalid with unsupported sandbox class "<value>", and placement of this Actor fails with WorkerPoolNotFound.
A missing SandboxConfig fails template preparation with SandboxConfig "<name>" not found.
A Worker image built for another class never registers, so the pool reports fewer ready Workers than its replica count.

Considerations before changing a running pool

Two constraints apply to a pool that already serves traffic.

Snapshots do not move between sandbox classes. An Actor cannot resume from a snapshot that a different class produced, so changing the class of a pool that holds live Actors strands their state.
Nodes fetch sandbox runtime assets from a public Google Cloud Storage URL, named in the SandboxConfig, independently of the storage backend that you configured for snapshots. A cluster with restricted egress needs a path to this URL, or a SandboxConfig that points at a copy you host.

Replacing workerImage rolls the Worker pods through the generated Deployment. The pods take a one-hour termination grace period, so an in-flight turn finishes rather than being cut off, and a rollout of a busy pool takes as long as its longest running turn.

Operational considerations Debug