Cloud Native Summit Kerala 2026

Kubernetes already knows
how to heal itself.

So why is your AI agent ignoring it?

Sat, August 22, 2026 · Le Meridien Kochi

Speaking today

RaghuReddy

Staff Platform Engineer (Security) @ Calix

Cloud Native Hyderabad · Organizer

Platform Engineering Security DevOps Speaker
raghureddycloud @raghureddycloud
Building secure platforms
The story

Live. Die.
Repeat.

Edge of Tomorrow, 2014

Pod dies. K8s restarts.
Problem remains.

Your cluster, this afternoon

Dying more was never the strategy.

He got out the moment somebody handed him the one thing the loop could never produce on its own: where the target actually was. New information, not more attempts.

The war story
Live. Die. Repeat.

Perfect recovery. Zero progress.

We built the loop again, by hand.
2,000 linesits own repo its own testsits own on-call
Resets
000
Perfect recovery, every single time.
Still on the same beach.
Restarting is not remediation
slides/assets/EdgeOfTomorrow_Second.png
The bridge

The reconcile loop
is the answer.

Observe. Diff. Act.

The problem

What everyone is building
right now.

class PodHealer:    def __init__(self):        self.state   = {}     # pending | retrying | failed | healed        self.backoff = {}        self.max_retries = 5    async def watch(self):        while True:            for pod in self.scrape_pods():        # regex over kubectl                if self.looks_unhealthy(pod):                    await self.maybe_retry(pod)            await asyncio.sleep(30)

Two thousand lines. Its own repo. Its own tests. Its own on-call rotation.

The title's actual question

Why is your agent
ignoring it?

What already ships

kube-controller-manager

OBSERVE list / watch the API DIFF desired vs actual ACT reconcile forever, without being asked
informerswork queues exponential backoffleader election idempotent reconcilesa decade in production

Code you do not write, do not test, and do not carry a pager for.

The mapping

The same loop, twice.

kube-controller-manager Observe list / watch the API Diff desired vs actual state Act reconcile your AI agent Observe list / watch the API Observe what is happening Diff desired vs actual state Reason what is wrong Act reconcile Act propose the fix

Not a metaphor. The same pattern with a different executor.

The signal layer

Stop scraping. Start asking.

Agent MCP tools read-only Kubernetes API readiness = what · events = why · logs = the string grep kubectl output no schema, no correlation
list_podsget_pod_status get_eventsget_pod_logs get_deploymentlist_secret_keys

The readiness probe says what is unhealthy. The events API says why.

MCP is what turns that into a tool call instead of a parser. That part is genuinely new. · k8sgpt already does diagnosis well; the argument here is about where the signal comes from.

Demotime
Demo · recorded, so you can run it yourself
Demo recording : online/offline
slides/assets/demo.mp4
The only sane default

The prescription,
not the surgery.

You do not need a new orchestration framework.
You need to stop ignoring the substrate.

Appendix · Q&A

"Why not let it write?"

Appendix · Q&A

Cost, and does it port?

Thank you, Kerala

RaghuReddy

Staff Platform Engineer (Security) @ Calix

Cloud Native Hyderabad · Organizer

raghureddycloud @raghureddycloud
Building secure platforms
0:00