Solo.io

Agents in kagent call tools from any MCP server, declared on the agent as a remote MCP server with the exact tool names it may use, and the built in tool server can apply manifests, patch and delete resources in the cluster as well as read pods, logs and events. kmcp scaffolds, builds and deploys new MCP servers in Python, Go, Java or TypeScript as Kubernetes resources, with secrets synced from a local env file. agentgateway federates many MCP servers behind one endpoint with OAuth, so each agent reaches outside systems under one access policy.

Sourcekagent.dev/docs/kagent/0.x/examples/human-in-the-loopread 2026-11-09

One kagent agent can call another as a tool: a coordinator lists a PromQL agent, or any other Agent resource in the same namespace or another, under its own tools and hands it work in the middle of its tool loop.

Calls to a subagent share one session by default for continuity, and an isolateSessions setting gives each call a fresh session so a coordinator can fan the same worker out in parallel.

Every agent also exposes an A2A endpoint for outside callers, teams can run LangGraph, CrewAI or Google ADK flows under kagent's governance, and an Agent Harness keeps a long running remote environment alive for extended work.

Sourcekagent.dev/docs/kagent/0.x/concepts/agents.mdread 2026-11-09

Grounding an agent in documents is a pattern rather than a managed service. A kagent guide crawls a documentation website, stores the content in a vector store and gives an agent a tool to query it, and agentgateway labs show retrieval and semantic caching at the gateway. Skills add written guidance agents load as needed. There is no built in knowledge base with connectors to wikis, drives or ticket systems, so a team assembles retrieval from MCP tools and its own vector database.

Sourcekagent.dev/docs/kagent/0.x/llms.txtread 2026-11-09

Any tool listed under requireApproval on an agent pauses the run before it executes, showing the proposed call, such as the YAML about to be applied, with Approve and Reject buttons in the dashboard. A rejection carries the person's stated reason back to the model so it changes course, read only tools run freely, and every agent has a built in ask_user tool for clarifying questions. Solo Enterprise adds approval workflows, review checkpoints and escalation paths, and agentgateway applies prompt guards to model traffic.

Sourcekagent.dev/docs/kagent/0.x/examples/human-in-the-loopread 2026-11-09

Solo Enterprise for kagent gives every agent a verifiable identity and governs each action by access policy, with on behalf of semantics and token exchange so an agent acts with a user's rights, plus RBAC, OIDC and OAuth sign in and mTLS between agents and tools. MCP servers can be secured with OAuth at agentgateway, and agentregistry holds an approved catalog with provenance checks. Solo's trust center lists SOC 2 Type 1 and Type 2 reports, redacted penetration test summaries, a CAIQ Lite questionnaire and its security policies.

Sourcetrust.solo.ioread 2026-11-09

Every agent action, tool call and decision path is traced with OpenTelemetry, which kagent starts and propagates on its own, so a run can be followed end to end in Jaeger or another tracing backend. Prompt auditing records every input and reply from each agent, sessions can be listed by ID from the CLI and opened in the dashboard, and the enterprise edition adds visual tracing of agent behavior. agentgateway adds end to end tracing of model and MCP traffic across teams, so one view covers the agent, its tools and its model calls.

Sourcekagent.dev/docs/kagent/0.x/llms.txtread 2026-11-09

Agent memory saves user intent, key learnings and preferences every fifth message as vectors in the kagent Postgres database, with timestamps and the source session, and pulls the closest matches into context before each reply. Memories are scoped to an agent and a user, expire after 15 days by default with ttlDays to change it, and can be listed, searched or deleted for one agent and user through the API or viewed in the dashboard. Agents also get save_memory and load_memory tools, and the embedding model is chosen separately from the main model.

Sourcekagent.dev/docs/kagent/0.x/concepts/agent-memoryread 2026-11-09

kagent, agentgateway and agentregistry install with Helm into the customer's own Kubernetes clusters, in any cloud, region or data center, and the enterprise editions add multicluster operation with high availability. Agent sessions, memories and configuration sit in a Postgres database the customer can point at its own instance, and model calls go to whichever provider and region the customer configures, including local models through Ollama. Solo does not host the runtime, so the buyer decides where every piece of agent data lives.

Sourcekagent.dev/docs/kagent/0.x/concepts/agent-memoryread 2026-11-09

kagent's public Agents Registry carries ready made agents a team installs and runs as they are or adapts: a Kubernetes agent, a Helm agent, an Istio agent, a kgateway agent, an observability agent that queries Prometheus and builds Grafana dashboards, a PromQL agent, a Cilium network policy agent and an Argo Rollouts converter, beside a catalog of MCP enabled tools. agentregistry gives an organization its own catalog of approved MCP servers, agents, skills and prompts to publish and reuse, and the kagent CLI scaffolds new agent projects for several frameworks.

Sourcekagent.dev/agentsread 2026-11-09

Work reaches a kagent agent when someone or something calls it: a chat in the dashboard, the kagent invoke command, an A2A client, the MCP server that exposes running agents, or a bot in Slack, Discord or Telegram connected over A2A. That covers the places people already talk to their infrastructure. No schedule, cron resource, webhook listener or cluster event watcher starts an agent on its own, so a recurring health check or an alert driven run needs an outside scheduler or alerting system to call the agent.

Sourcekagent.dev/docs/kagent/0.x/llms.txtread 2026-11-09

Each kagent agent points at a ModelConfig the team defines, with routes to OpenAI, Anthropic, Gemini, Google Vertex AI, Azure OpenAI, Azure AI Foundry, Amazon Bedrock, Ollama, SAP AI Core, xAI or any OpenAI compatible endpoint, and memory can use a different embedding model. Requests can also run through agentgateway, which adds model aliasing, token weighted rate limits, failover between providers and routing rules, so a platform team can switch a model for every agent without editing them.

Sourcekagent.dev/docs/kagent/0.x/llms.txtread 2026-11-09

Agents, tools, model configurations and MCP servers are Kubernetes custom resources with a published API reference, so teams manage them with kubectl, Helm or GitOps. The kagent CLI installs, initializes, builds, deploys and invokes agents and MCP servers, REST endpoints manage sessions and memories, and the control plane runs an MCP server that exposes every running agent to outside clients. Every agent also implements A2A, so other agents and applications can call it directly.

Sourcekagent.dev/docs/kagent/0.x/llms.txtread 2026-11-09

kagent has no evaluation harness, scored test set or quality gate in its documentation. Local development runs an agent project in Docker Compose for an interactive chat before deploying it, and the debug guide covers agents that fail to start or answer. Solo's marketing graphics list evaluation among agent capabilities and a blog post describes an agentevals benchmarking project, but no product page or guide shows how a team scores an agent's answers or blocks a release on them.

Sourcekagent.dev/docs/kagent/0.x/llms.txtread 2026-11-09

kagent agents do not drive a browser, desktop or remote computer. They act through MCP tools and Kubernetes APIs, and Agent Substrate runs agents and stateful workloads in sandboxed pods, which isolates execution without operating a screen. agentdesktop watches and secures the agents employees run on their own machines, giving visibility and control over them rather than letting an agent click through desktop applications.

Sourcesolo.ioread 2026-11-09