Guides

One task per page. The build guides cover the agent and its guardrails; the operate guides take it to a cluster and keep it healthy.

Build

• Develop your agent

  The generated project, the service and its endpoints, tools and prompts, answers in a JSON shape you declare, the fastapi and langgraph-server runtimes, run and playground.

• Authentication

  One policy on every surface: shared-bearer, per-user jwt (with local dev tokens) or a custom policy of your own.

• Outbound API policy

  Declare the APIs tools may call in api-policy.yaml, widen or narrow access with graph-agents-cli api, and let lint check every declared call.

• Human approval

  Make chosen calls wait for the requester or a second person, decide them with approvals, and understand what an approval binds.

• Agents calling agents

  Let an agent ask other agents for its user: peer add, token exchange, relayed approvals, the system view for many projects, and the threat model.

• Evaluation

  Datasets, deterministic checks, judges and quality metrics, and the gate that eval run enforces locally and in CI.

• Extensions

  Override or add CLI commands from an extension repository or a local path (experimental).

Operate

• Deploy to Kubernetes

  Environments, the Helm chart, local clusters, direct deploys, rollouts, rollback and an external database.

• CI/CD

  The skip, helm-push and argocd modes, the generated workflows and the GitHub settings they need.

• Secrets

  The allow-listed app Secret: apply, check and rotate keys without printing a value.

• Observability

  JSON logs, Prometheus metrics, run records and opt-in tracing with LangSmith or OTLP.

• Upgrading projects

  scaffold upgrade with a 3-way merge against the build that created the project, and migrating from 0.1.0.

• Offline profile

  Run the whole lifecycle without internet access: on-network models, mirrors and checks.

• Security & production

  The security model, what it leaves to you, and the checklist before production traffic.