Deploy an agent

Goal: get your own agent image running as a governed, autoscaled service.

Prerequisites: the platform installed; a namespace you can write to; a model available via a ModelRoute; your agent packaged on a supported base image.

1. Describe the agent

An agent is one resource. The minimum is an image and an execution model:

apiVersion: agents.ctxmesh.ai/v1beta1
kind: AgentDeployment
metadata:
  name: support-agent
  namespace: my-team
spec:
  image: ghcr.io/my-org/support-agent:1.0.0
  executionModel: serving        # serving | eventing | job
  # The model is chosen in your code by calling MODEL_GATEWAY_URL with model="<ModelRoute name>".
  scaling:
    min: 0                        # scale to zero when idle (default)
    max: 3

Apply it:

kubectl apply -f support-agent.yaml

2. Watch it come up

kubectl get agentdeployment support-agent -n my-team -w
# Ready transitions to True once the serving revision is up.
kubectl get agentdeployment support-agent -n my-team \
  -o jsonpath='{.status.conditions[?(@.type=="Ready")].status} {.status.url}{"\n"}'

The controller injects the gateway URL + launcher config, wires tracing, and reports readiness and the serving URL on status.

3. Talk to it

Send a request to status.url (or use the console Playground). Every turn is traced — open the agent’s runs in the console to see the step → tool → model tree, cost, and any guardrail decisions.

4. Attach governance (by reference)

Governance is opt-in and reusable. Author the policies once (see the linked guides) and reference them:

spec:
  image: ghcr.io/my-org/support-agent:1.0.0
  executionModel: serving
  guardrailPolicyRef: default-guardrails   # content rules — /guides/guardrails/
  approvalPolicyRef: sensitive-tools       # human approval — /guides/approvals/
  evalSuiteRef: support-quality            # gate releases — /guides/evals-and-the-deploy-gate/
  feedbackStoreRef: support-feedback       # feedback model — /guides/feedback-and-improvement/
  sessionMemory:
    scope: session                         # per-conversation memory — /guides/memory-and-sessions/

A dangling reference fails closed: the agent goes Ready=False and is held rather than served ungoverned.

When to use / when not

• serving for interactive request/response agents; eventing for async/event-driven work; job for batch. See Execution models.
• Leave scaling.min: 0 for scale-to-zero unless you need warm capacity (cold-start latency vs cost).

Defaults

• executionModel defaults to serving; scaling defaults to min: 0, max: 3.
• No evalSuiteRef ⇒ no deploy gate (zero overhead). No policy refs ⇒ today’s ungoverned behavior.

Failure modes

• Image pull / crash → the revision is not Ready; kubectl describe + the pod logs show why.
• Dangling policy ref → Ready=False with the reason on the condition (fix or remove the ref).
• Model alias unresolved (no matching ModelRoute) → model calls fail fast at the gateway.

Next

Connect a model provider · Guardrails · Gate a release on an eval suite · AgentDeployment reference