Change anything on its own
The AI agent holds no credential that can write. Every change is made by a human. Structural, not a setting.
Kubernetes operations
We run your clusters. An AI watches. Humans take action.
opsturbo puts a harness on your fleet's event stream: an AI agent that reads every event, investigates, and dispatches only what matters to a human. It can never write to your cluster.
02:14eventcheckout-7f2 · OOMKilled
02:14agentbatch job hit its memory limit
02:15—nothing to report
02:41eventnode 10-0-3-7 · NotReady
02:41agent9 pods stuck · checkout losing capacity
02:42agentpaged on-call · suggest: cordon + drain
02:51humandrained the node · checkout failed over
03:02—cleared
Where "the AI fixed it" is not an acceptable sentence
- fintech
- Every change is a human's, and traceable.
- healthcare
- Change is controlled. An observer never widens it.
- insurtech
- Quiet clusters. Silence is checked, not assumed.
- ad tech
- Spikes by the minute. Nobody woken for the ones that pass.
How a night goes
Seven steps. One belongs to a human.
Watch
Every event, every cluster.
Investigate
Describe, logs, pressure.
Decide
Noise, or not. Mostly noise.
Page
An answer, not an alert. Evidence and a suggestion.
Fix
A human. On their credential.
Learn
Every run leaves a sealed transcript. Replayable, reviewable.
Report
Monthly. Hours saved, pages avoided.
What we never do
Move your data
Read where it happens. Only sealed findings travel.
Hold your keys
Model, cluster, transcript — yours, on your hosts.
- 0
credentials that can write
- 8
tools it may use alone — none mutate
- 30
model turns, max, per hour
- 50
events batched into one look
- 1
human in every fix
Not settings. Constants, compiled into the binary. Config can narrow them — never widen.
Why the guarantee holds
Policy is code, not configuration.
A guard runs before every tool, not after. A tool not on the list is denied before it executes, and the refusal comes back as an ordinary result. A prompt injection meets code, not instructions.
read the event streamallow
read the cluster · never secretsallow
check a deliveryallow
check its own usageallow
read a runbookallow
speak to the humanallow
send anything outa human says yes
delete, restart, drain — anything elsedenied
Two ways to work with us
Same system. The difference is who holds the pager.
Managed
- pager
- ours
- runbooks
- we write, you approve
- price
- per cluster, per month
Self-run
- pager
- yours
- runbooks
- yours, we review
- price
- less, with support
How we work
You get senior attention where it counts, and quiet everywhere else.
The AI agent does the watching, the reading and the first decision. A human is spent only where something critical changes.
watchagent
investigateagent
decideagent
fixhuman
learnagent
reportagent
Run clusters? Put read-only agents on them.
opsturbo is early. Tell us about your fleet and we will tell you when a harness is ready for it.
early-access@opsturbo.com