Release Notes#

The kagent documentation shows information only for the latest release. If you run an older version, review the release notes to understand the main changes from version to version.

For more details on the changes between versions, review the kagent GitHub releases.

v0.8#

Review this summary of significant changes from kagent version 0.7 to v0.8.

• Human-in-the-Loop (HITL) — tool approval gates and interactive ask_user tool.
• Agent Memory — vector-backed long-term memory for agents.
• Go ADK runtime — new Go-based agent runtime for faster startup and lower resource usage.
• Agents as MCP servers — expose A2A agents via MCP for cross-tool interoperability.
• Skills — markdown knowledge documents loaded from OCI images or Git repositories.
• Go workspace restructure — the Go codebase is split into api, core, and adk modules for composability.
• Prompt templates — reusable prompt fragments from ConfigMaps using Go template syntax.
• Context management — automatic event compaction for long conversations.
• AWS Bedrock support — new model provider for AWS Bedrock.
• PostgreSQL-only database backend — SQLite support has been removed. PostgreSQL is now the only supported database backend.

Human-in-the-Loop (HITL)#

You can now use two Human-in-the-Loop mechanisms that can pause agent execution and wait for user input.

Tool Approval — You can mark specific tools as requiring user confirmation before execution by using the requireApproval field in the Agent CR. When the agent calls a tool that requires approval, the UI presents Approve/Reject buttons. The reason provided for rejection gets used as context for the LLM.

Ask User — A built-in ask_user tool is automatically added to every agent. Agents can pose questions to users with predefined choices (single-select, multi-select) or free-text input during execution.

For more information, see the Human-in-the-Loop example and the blog post.

Agent Memory#

Your agents can now automatically save and retrieve relevant context across conversations using vector similarity search. Memory is built on the Google ADK memory implementation and uses the same kagent database (PostgreSQL).

When you enable memory on an agent, it receives three additional tools: save_memory, load_memory, and prefetch_memory. Every 5th user message, the agent automatically extracts key information, such as user intent, key learnings, preferences.

You can configure memory in the Agent CR or through the UI when you create or edit an agent by selecting an embedding model and TTL.

For more information, see Agent Memory.

Go ADK Runtime#

You can now choose between two Agent Development Kit runtimes: Python (default) and Go. The Go ADK provides significantly faster startup (~2 seconds vs ~15 seconds for Python) and lower resource consumption.

Select the runtime via the runtime field in the declarative agent spec.

spec:
  type: Declarative
  declarative:
    runtime: go

The Go ADK includes built-in tools: SkillsTool, BashTool, ReadFile, WriteFile, and EditFile.

For more information, see Agents and the blog post.

Agents as MCP Servers#

Agent-to-Agent (A2A) agents are now exposed as MCP servers via the kagent controller HTTP server. This enables cross-tool interoperability — any MCP-compatible client can consume agents, not just the A2A protocol.

Skills#

Your agents can now load markdown-based knowledge documents (skills) that provide domain-specific instructions, best practices, and procedures. Skills load at agent startup and are discoverable through the built-in SkillsTool.

You can load skills from two sources.

• OCI images. Container images containing skill files.
• Git repositories. Clone skills directly from Git repos, with support for private repos via HTTPS token or SSH key authentication.

For more information, see Agents.

Go Workspace Restructure#

The Go code now uses a Go workspace with three modules: api, core, and adk. This makes the codebase more composable for you if you want to pull in parts of kagent (such as the API types or ADK) without importing all dependencies.

Prompt Templates#

Agent system messages now support Go text/template syntax. You can store common prompt fragments, such as safety guardrails or tool usage best practices, in ConfigMaps and reference them with {{include "alias/key"}} syntax.

The kagent-builtin-prompts ConfigMap ships with five reusable templates: skills-usage, tool-usage-best-practices, safety-guardrails, kubernetes-context, and a2a-communication.

For more information, see Agents.

Context Management#

Long conversations can now be automatically compacted to stay within LLM context windows. You can configure the context.compaction field to enable periodic summarization of older events while preserving key information.

For more information, see Agents.

AWS Bedrock Support#

You can now use AWS Bedrock as a model provider, allowing your agents to use Bedrock-hosted models.

PostgreSQL-Only Database Backend#

SQLite support has been removed from kagent. PostgreSQL is now the only supported database backend.

What changed:

• The database.type configuration option is removed.
• SQLite-related Helm values (database.sqlite.*) are removed.
• A bundled PostgreSQL instance is deployed by default via database.postgres.bundled.enabled: true. The bundled image is postgres:18 (standard PostgreSQL without pgvector).
• database.postgres.vectorEnabled now defaults to false. Set it to true only when using a PostgreSQL server that has the pgvector extension installed.
• database.postgres.bundled.enabled and url/urlFile are now independent controls. You can keep the bundled pod running while pointing the controller at an external database, which is useful for migration.
• The bundled instance database name, username, and password are hardcoded to kagent. Credentials are stored in a Kubernetes Secret instead of a ConfigMap.
• The database.postgres.bundled.database, bundled.user, and bundled.password configuration options are removed.

Why this change:

• SQLite lacks pgvector support, requiring separate code paths for memory and vector search.
• SQLite's single-writer constraint prevents horizontal scaling of the controller.
• Divergent SQL dialects between SQLite and PostgreSQL required maintaining duplicate code paths.
• PostgreSQL was already the recommended production backend.

Migration:

If you were using the default SQLite backend, no migration is needed. The bundled PostgreSQL is deployed automatically. You can optionally customize the bundled instance via database.postgres.bundled.* (storage size, image) as needed. See the Database configuration guide for details.

Note that for production deployments, use your own external PostgreSQL instance. If you already are, you can keep your database.postgres.url or database.postgres.urlFile settings as before. If your external PostgreSQL has the pgvector extension and you were using vector-based memory features, set database.postgres.vectorEnabled: true since the default has changed to false.

Additional Changes#

• API key passthrough for ModelConfig.
• Custom service account override in agent CRDs.
• Voice support for agents.
• UI dynamic provider model discovery for easier model configuration.
• CLI --token flag for kagent invoke API key passthrough.
• CVE fixes across Go, Python, and container images.

v0.7#

Review the main changes from kagent version 0.6 to v0.7, then continue reading for more detailed information.

• kmcp is installed by default when you install kagent
• New feature to develop agents locally without a Kubernetes cluster
• New kgateway.dev/discovery label
• Installation profiles

kmcp installed by default#

Now, kmcp is installed automatically with kagent, so you can use kmcp functionality out of the box.

This change is enabled by the new default values of kmcp.enabled=true in both the kagent and kagent-crds Helm charts.

Existing kmcp installations#

If you already have kmcp installed separately, upgrade your existing Helm releases with the kmcp.enabled=false flag set for both the kagent and kagent-crds charts.

Example commands:

kagent-crds Helm release:

helm upgrade --install kagent-crds oci://ghcr.io/kagent-dev/kagent/helm/kagent-crds \
  --namespace kagent \
  --set kmcp.enabled=false

kagent Helm release:

helm upgrade --install kagent oci://ghcr.io/kagent-dev/kagent/helm/kagent \
  --namespace kagent \
  --set kmcp.enabled=false

Local agent development#

Develop and test agents locally on your machine without needing a Kubernetes cluster. As part of this feature, the kagent CLI includes new commands to scaffold, build, run, and deploy agents.

For more information, see the local development guide.

Discovery label#

Now, you can add a discovery label to MCPServer kmcp resources. By default, discovery is enabled.

If you plan to use your kmcp resources later with kagent and agentgateway, add the kagent.dev/discovery=disabled label to your MCPServer resource. Then, kagent does not automatically discover MCP servers. This way, you can have agentgateway in front of your kmcp servers so that the agent-tool traffic is routed correctly through agentgateway.

Installation profiles#

By default, kagent installs a demo profile with agents and MCP tools preloaded for you. If you don't want these default agents, you can disable them with the minimal profile.

For the CLI: kagent install --profile minimal

For Helm installations: Individually disable the default agents with Helm values or --set flags, such as --set agents.argo-rollouts-agent.enabled=false. You can also use Helm to update the resource limits and requests for each agent.

v0.6#

Review the main changes from kagent version 0.5 to v0.6, then continue reading for more detailed information.

• The apiVersion field in the kagent CRDs is now kagent.dev/v1alpha2.
• A new Helm chart for kmcp CRDs is available.
• API string references to resources in other namespaces in the format namespace/name now fail. Instead, the APIs have a separate field for you to specify the namespace of the resource.
• The Tools API moves or eliminates some APIs entirely in favor of new kmcp APIs.
• The Agents APIs now require a top-level type field to support the new BYO agent functionality.
• The ModelConfig APIs rename the secret name field from apiKeySecretRef to apiKeySecret.
• Memory APIs are not supported in ADK.

Upgraded API version#

The apiVersion field in the kagent CRDs is now kagent.dev/v1alpha2.

New! Helm chart for kmcp CRDs#

Previously, the kagent installation included only one CRD Helm chart. As of v0.6.3, the MCPServer CRD is part of a separate kmcp Helm subchart. This kmcp subchart is installed for you when you install the kagent CRD Helm chart.

1. If you installed the separate kmcp CRD Helm in earlier versions of v0.6, uninstall the Helm chart.

   helm uninstall kmcp-crds -n kagent

2. Install the kagent CRD Helm chart that includes the kmcp subchart.

   helm install kagent-crds oci://ghcr.io/kagent-dev/kagent/helm/kagent-crds \
     --namespace kagent \
     --create-namespace

General changes#

namespace/name references: API string references to resources in other namespaces in the format namespace/name now fail. Instead, the APIs have a separate field for you to specify the namespace of the resource.

Local development buildx access: The make helm-install command now creates a local Docker registry to push development images to. As part of the build process, you might need to allow the buildx builder to access your host network. For more information, see the developer docs in the kagent repo.

Tools APIs#

The Tools-related APIs are split up into several different APIs. Some functionality is moved to kmcp, such as the ToolServer API.

ToolServer#

The ToolServer API is completely removed from kagent. Instead, use other resources including some kmcp APIs to create and manage tools.

Stdio ToolServer now in kmcp MCPServer#

Flip through the following tabs to understand the API differences between the old kagent ToolServer and the new method of using kmcp along with a kagent MCPServer and Kubernetes Service for the Stdio transport type.

apiVersion: kagent.dev/v1alpha2
kind: ToolServer
metadata:
  name: mcp-grafana
  namespace: kagent
spec:
  config:
    stdio:
      command: /app/python/bin/mcp-grafana
      args:
        - -t
        - stdio
        - debug
      readTimeoutSeconds: 30
      envFrom:
      - name: "GRAFANA_URL"
        value: my-url.com
      - name: "GRAFANA_API_KEY"
        valueFrom:
          type: Secret
          key: "grafana"
          valueRef: kagent-toolserver-secret
  description: ""

apiVersion: kagent.dev/v1alpha1
kind: MCPServer
metadata:
  name: grafana
spec:
  deployment:
    image: "mcp/grafana:latest"
    port: 3000
    cmd: "/app/mcp-grafana"
    args:
      - "--transport"
      - "stdio"
    env:
      GRAFANA_URL: my-url.com
    secretRefs:
      - name: grafana-api-key
  transportType: "stdio"

apiVersion: v1
kind: Secret
metadata:
  name: grafana-api-key
type: Opaque
data:
  GRAFANA_API_KEY: my-base-64-ikey

HTTP ToolServer moved to RemoteMCPServer#

ToolServer resources that used type: streamableHttp are now configured as RemoteMCPServer resources. For more detailed information, review the API definitions:

• ToolServer: toolserver_types.go
• RemoteMCPServer: remotemcpserver_types.go
• MCPServer: mcpserver_types.go

apiVersion: kagent.dev/v1alpha2
kind: ToolServer
metadata:
  name: kagent-tool-server
spec:
  config:
    type: streamableHttp
    streamableHttp:
      url: "http://kagent-tools.kagent:8084/mcp"
      timeout: 30s
      sseReadTimeout: 5m0s
  description: "Official kagent tool server"

apiVersion: kagent.dev/v1alpha2
kind: RemoteMCPServer
metadata:
  name: kagent-tool-server
spec:
  url: "http://kagent-tools.kagent:8084/mcp"
  timeout: 30s
  sseReadTimeout: 5m0s
  description: "Official kagent tool server"

Kubernetes Services as HTTP MCP servers#

Now, you can use Kubernetes Services as MCP Servers.

apiVersion: v1
kind: Service
metadata:
  name: kagent-querydoc
  namespace: kagent
spec:
  ports:
    - name: http
      port: 8080
      protocol: TCP
      targetPort: http
---
apiVersion: kagent.dev/v1alpha2
kind: ToolServer
metadata:
  name: kagent-querydoc
  namespace: kagent
spec:
  description: Queries a documentation site
  config:
    sse:
      url: http://kagent-querydoc.kagent.svc.cluster.local/sse

apiVersion: v1
kind: Service
metadata:
  labels:
    kagent.dev/mcp-service: "true"
  annotations:
    kagent.dev/mcp-service-path: /sse
    kagent.dev/mcp-service-port: 8080
    kagent.dev/mcp-service-protocol: SSE
  name: kagent-querydoc
  namespace: kagent
spec:
  ports:
  - appProtocol: mcp
    name: http
    port: 8080
    protocol: TCP
    targetPort: http

Agent APIs#

API to specify the MCP server#

tools:
  - type: McpServer
    mcpServer:
      toolServer: kagent-querydoc
      toolNames:
        - query_documentation

tools:
  - type: McpServer
    mcpServer:
      name: kagent-querydoc
      kind: Service
      toolNames:
        - query_documentation

The toolServer field has been removed, and has been replaced with the following: name kind apiGroup This allows for specifying the 3 different types which may be used. Those are: RemoteMCPServer Service MCPServer Here are the 2 api defs:

• v1alpha1
• v1alpha2

Top-level field for Agent APIs#

A new top-level type field is added to the Agents API. For existing Agents, set the type to Declarative, and then nest the previous Agent configuration inline under the declarative setting.

This change supports the new type for BYO agents.

apiVersion: kagent.dev/v1alpha2
kind: Agent
metadata:
  name: k8s-agent
  namespace: {{ include "kagent.namespace" . }}
  labels:
    {{- include "kagent.labels" . | nindent 4 }}
spec:
  description: An Kubernetes Expert AI Agent specializing in cluster operations, troubleshooting, and maintenance.
  systemMessage: |
    # Kubernetes AI Agent System Prompt

    You are KubeAssist, an advanced AI agent
    # ... (truncated for brevity)

apiVersion: kagent.dev/v1alpha2
kind: Agent
metadata:
  name: k8s-agent
  namespace: kagent
spec:
  description: An Kubernetes Expert AI Agent specializing in cluster operations, troubleshooting, and maintenance.
  type: Declarative
  declarative:
    systemMessage: |
      # Kubernetes AI Agent System Prompt

      You are KubeAssist, an advanced AI agent
      # ... (truncated for brevity)

New! BYO agents#

A new agent type has been added to the Agents API so that you can bring your own (BYO) agent. The agent must be written in ADK, with other frameworks under development.

BYO Agent example configuration. For more information, see the BYO Agent guide.

apiVersion: kagent.dev/v1alpha2
kind: Agent
metadata:
  name: basic-agent
  namespace: kagent
spec:
  description: This agent can do anything.
  type: BYO
  byo:
    deployment:
      image: my-byo:latest
      env:
        - name: GOOGLE_API_KEY
          valueFrom:
            secretKeyRef:
              name: kagent-google
              key: GOOGLE_API_KEY

ModelConfig API#

The secret name field is renamed from apiKeySecretRef to apiKeySecret.

ModelInfo removed#

The modelInfo setting is removed from the ModelConfig API.

Supported LLM providers are pre-configured by the kagent-dev/autogen project fork for you by default. Trying to override these default settings, such as to enable vision for image recognition, could cause unexpected behavior in models that do not support these settings. Therefore, the modelInfo field is removed.

Memory API#

The Memory API is not supported in ADK. The agent development kit is required to bring your own agents. As such, the Memory docs are removed.