kagent v0.9.1 — What's New in Kubernetes-Native Agentic AI#

Published May 2026 | kagent-dev/kagent v0.9.1

kagent has been making serious waves in the Kubernetes-native AI space. The project — a framework for building cloud-native agentic AI on top of Kubernetes — just shipped v0.9.1, and it's packed with features that make running AI agents in production significantly more practical.

If you've been following the cloud-native AI conversation, you know the gap between "demo on my laptop" and "running agents reliably in Kubernetes" is huge. kagent bridges that gap. And v0.9.1 narrows it even further.

Here's what's new.

1. AWS Bedrock Support for Claude Models with Extended Thinking#

The most impactful feature in this release: full support for Claude models on AWS Bedrock with top_k sampling and extended thinking modes.

What this means:

• You can now deploy Claude models through kagent on AWS Bedrock with fine-grained control over output quality (top_k parameter) and reasoning depth (extended thinking)
• For complex reasoning tasks, extended thinking lets Claude models spend more compute on the problem — think of it as giving the model "more time to think" before responding
• This is a game-changer for teams running AI agents on AWS who need the reliability and compliance posture that Bedrock provides

The feature comes from contributor mesutoezdil and is documented in PR #1732.

2. Secret Creation via REST API with ModelConfig#

Managing model secrets used to be a CLI-only or manual process. v0.9.1 changes that with secret creation directly through the REST API.

Now you can include secrets in your ModelConfig resource when making API calls. This means:

• Fully programmatic secret management for model configurations
• Seamless integration with CI/CD pipelines — no more manual kubectl apply for secrets
• Better automation for multi-tenant agent deployments

Again, credit to supreme-gg-gg via PR #1781.

3. Sidecar Container Support#

Kubernetes agents don't always run in isolation — you often need log shippers, service mesh proxies, or custom monitoring agents running alongside your AI agent.

v0.9.1 introduces extraContainers support in SharedDeploymentSpec, letting you attach sidecar containers to agent deployments. This is essential for:

• Enterprise observability (collecting agent metrics and logs alongside the AI workload)
• Service mesh integration (linkerd, istio sidecars)
• Custom data processing pipelines running in the same pod

PR #1724 by mesutoezdil.

4. Git-Based Skill Configuration via UI#

Skills are a core concept in kagent — they're the tools and capabilities your AI agents use. Previously, configuring skills from external repositories required YAML manipulation.

v0.9.1 adds a new UI for configuring skills directly from Git repositories. This makes the workflow:

• Visual and self-serve — no more fighting with manifests
• Developer-friendly — point to a Git repo, configure the skill, done
• GitOps-compatible — skills defined in code, version-controlled, auditable

Built by peterj in PR #1747, with a UI cleanup follow-up in #1764.

5. Helm Chart — Configurable Agent Runtime#

The kagent Helm chart got a significant upgrade: you can now configure the agent runtime directly through Helm values, without patching manifests or writing custom overlays.

This matters because:

• Simplifies Helm-only deployments (no need for kustomize or manual patches)
• Makes it easier to parameterize runtime settings across environments
• Reduces the operational overhead of managing kagent at scale

Contributed by marcofranssen — and notably, this was his first contribution to the project (PR #1697).

Also Notable: Bug Fixes & Improvements#

Beyond the headline features, v0.9.1 includes a solid batch of bug fixes:

• Bubblewrap in CI — fixed bubblewrap so it works correctly in CI environments (PR #1736)
• Image scan fixes — resolved main image scan failures that were blocking security pipelines (PR #1742)
• CLI config defaults — the CLI now properly honors configuration defaults (PR #1746)
• MCP tool call headers — allowedHeaders from incoming requests are now correctly forwarded to MCP tool calls (PR #1733)
• OTLP exporter protocol — the Python OTLP exporter now respects the OTEL_EXPORTER_OTLP_PROTOCOL environment variable instead of hardcoding gRPC (PR #1681)
• Structured logging — replaced stdlib log with structured logr in the memory handler for better observability (PR #1773)

Community Spotlight#

What stands out about this release is the diverse and growing contributor base:

12 contributors total. That's a healthy pace for a project that's still relatively early in its lifecycle.

What's Next#

kagent is building the infrastructure layer for running AI agents at Kubernetes scale. With v0.9.1, the project is clearly moving from "experimental framework" toward "production-ready tooling."

The features in this release — sidecar support, Git-based skill management, Bedrock integration, and API-driven secret management — are all signals of a project maturing toward real enterprise adoption.

If you're running AI agents on Kubernetes and haven't looked at kagent yet, v0.9.1 is a good time to start.

Want to try it? Check out the kagent GitHub repo and join the Discord.