Agent to Agent tool reference created before the referenced Agent exists is never retried

[EdmondDantes21]

Prerequisites

• I have searched the existing issues to avoid creating a duplicate
• By submitting this issue, you agree to follow the Code of Conduct
• I am using the latest version of the software.t
• I can consistently reproduce this issue

Affected Service(s)

Controller Service

Impact/Severity

Blocker: a fresh multi-agent install (Helm/GitOps/Argo/any tool that doesn't strictly order same-kind CRDs) can silently and permanently leave an agent broken with no indication anything will ever fix it.

Bug Description

When an Agent/SandboxAgent CR references another Agent that doesn't exist yet at apply time, the referencing agent gets stuck at Accepted: False forever, even after the referenced agent is created. The controller never retries.

The only things that ever re-trigger reconciliation of the broken parent are editing its spec, changing its labels, or restarting the kagent-controller pod.

Steps To Reproduce

1. Create the parent agent, referencing a child that does not exist yet:

   apiVersion: kagent.dev/v1alpha2
   kind: Agent
   metadata:
     name: repro-parent-agent
     namespace: default
   spec:
     type: Declarative
     declarative:
       modelConfig: repro-placeholder-model-config
       runtime: python
       systemMessage: "test agent, never meant to run"
       tools:
       - type: Agent
         agent:
           kind: Agent
           name: repro-child-agent

   $ kubectl apply -f repro-parent.yaml
   agent.kagent.dev/repro-parent-agent created
   

2. Check its status (~5s later):

   $ kubectl get agent repro-parent-agent -n default -o jsonpath='{.status.conditions}'
   [{"lastTransitionTime":"2026-08-27T13:34:16Z",
     "message":"failed to compile agent default/repro-parent-agent: Agent.kagent.dev \"repro-child-agent\" not found",
     "observedGeneration":1,"reason":"ReconcileFailed","status":"False","type":"Accepted"}]
   

3. Create the referenced child agent:

   apiVersion: kagent.dev/v1alpha2
   kind: Agent
   metadata:
     name: repro-child-agent
     namespace: default
   spec:
     type: Declarative
     declarative:
       modelConfig: repro-placeholder-model-config
       runtime: python
       systemMessage: "test agent, never meant to run"
       tools: []

   $ kubectl apply -f repro-child.yaml
   agent.kagent.dev/repro-child-agent created
   

   (repro-child-agent now exists and is kubectl get-able immediately.)

4. Re-check the parent's status 60s, then ~3 more minutes later — untouched:

   $ kubectl get agent repro-parent-agent -n default -o jsonpath='{.status.conditions[0]}'
   {"lastTransitionTime":"2026-08-27T13:34:16Z",
    "message":"failed to compile agent default/repro-parent-agent: Agent.kagent.dev \"repro-child-agent\" not found",
    "observedGeneration":1,"reason":"ReconcileFailed","status":"False","type":"Accepted"}
   

   Identical lastTransitionTime and observedGeneration — never re-reconciled.

Expected Behavior

Creating repro-child-agent should, on its own, cause repro-parent-agent to be re-reconciled and re-evaluate the tool reference — the same way creating a ModelConfig, RemoteMCPServer, ConfigMap, or MCPServer that an agent depends on already does.

Actual Behavior

repro-parent-agent stays pinned at Accepted: False / ReconcileFailed with a stale "not found" message indefinitely, even though the resource it complains about has existed for minutes. Only an incidental spec/label change or a full controller pod restart unsticks it.

Environment

• Kubernetes: kind, v1.36.1
• kagent: v0.9.12
• Install method: helm

Are you willing to contribute?

• I am willing to submit a PR to fix this issue