Summary
kagent install with the Anthropic provider deploys claude-3-5-haiku-20241022 which Anthropic has retired. Every agent fails immediately with:
Error code: 404 - {'type': 'error', 'error': {'type': 'not_found_error', 'message': 'model: claude-3-5-haiku-20241022'}}
Root cause
The model is hardcoded in helm/kagent/values.yaml:
anthropic:
provider: Anthropic
model: "claude-3-5-haiku-20241022" # retired
Impact
A fresh kagent install --profile demo --provider anthropic is broken out of the box. The user sees no error during install — the failure only surfaces when they try to chat with an agent.
Two requested fixes
1. Update the default model to the current Haiku release:
anthropic:
model: "claude-haiku-4-5-20251001" # Haiku 4.5 — current
2. Add model validation at install time — before writing the ModelConfig into the cluster, make a lightweight API call to verify the model ID is active. This catches retired or mistyped model names immediately with a clear error, rather than silently deploying a broken configuration that only fails mid-chat.
Error: Anthropic model 'claude-3-5-haiku-20241022' is not available.
Did you mean 'claude-haiku-4-5-20251001'? Run 'kagent models list --provider anthropic' to see available models.
Workaround
kubectl patch modelconfig default-model-config -n kagent \
--type=merge \
-p "{\"spec\":{\"model\":\"claude-haiku-4-5-20251001\"}}"```
Summary
kagent installwith the Anthropic provider deploysclaude-3-5-haiku-20241022which Anthropic has retired. Every agent fails immediately with:Root cause
The model is hardcoded in
helm/kagent/values.yaml:Impact
A fresh
kagent install --profile demo --provider anthropicis broken out of the box. The user sees no error during install — the failure only surfaces when they try to chat with an agent.Two requested fixes
1. Update the default model to the current Haiku release:
2. Add model validation at install time — before writing the ModelConfig into the cluster, make a lightweight API call to verify the model ID is active. This catches retired or mistyped model names immediately with a clear error, rather than silently deploying a broken configuration that only fails mid-chat.
Workaround