Product1 publisher3 min readPublished
Akuity ties every agent action in its delivery pipeline to the person who connected it
Akuity's Agentic Control Plane authenticates each agent request as a named engineer and carries that person's existing permissions across every cluster it manages, with a policy layer above that can still block production actions.
The Product Desk · Product desk
What happened
- Akuity introduced Agentic Control Plane, a layer that lets AI agents read its pipeline data and act on it under the permissions the platform already enforces.
- Every agent request is routed through the control plane and authenticated as the specific person who connected the agent. No shared service account sits in the path.
- Akuity said the data agents reach is the pipeline's operational record: deployment history, cluster health and change lineage across a fleet, down to the commit behind a deployment.
- Michael Goodness, a principal DevOps engineer at Major League Baseball, said the control plane surfaced more than 100 degraded applications within hours and traced one project's systemic issue to root cause in under 10 minutes.
- Akuity was founded by the creators of Argo CD and Kargo, and Lead Edge Capital led a $20 million Series A for the company in 2022.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision Teams whose agents already reach production through open-source MCP servers now have a named alternative to weigh, and the price of an audit trail is routing delivery through one vendor's control plane.
- constraint Scoping the agent to the connecting human caps what it can ever do at what that person can do, so the widest-privileged engineer who plugs in an agent sets the widest reach, and the policy layer can only narrow it.
- exposure One engineer's account becomes the record auditors follow for whatever the agent does at 3am, including actions nobody was awake to approve.
An engineer types a promote command into a chat window and the release moves. Akuity said the same approval gates apply whether a person or an agent made the request, and that an agent pointed at an empty repository can stand up the environments and promotion rules for a new service in an afternoon [8].
The authentication design underneath is the part a platform team has to sign off. Akuity said many agents already reach production through open-source MCP servers, which hand over raw access with no identity carried through and no audit trail of what the agent touched [3]. Its own MCP Server connects agents to the platform over the Model Context Protocol [2]. The control plane carries the connecting engineer's existing permissions across every instance and cluster Akuity manages, so there are no separate credentials to issue or rotate [5]. A policy layer above it can block sensitive production actions that user is otherwise cleared to perform [6].
That makes one person the ceiling. An agent connected by an engineer with broad production rights inherits broad production rights, and the policy layer can only subtract from them [22]. Akuity's announcement does not describe what happens to the connection when that person is deprovisioned [23].
What a customer has described so far sits on the read side. "That's exactly the workflow we wanted: Identify the problem, tell us if it's systemic and hand us the fix," Goodness said [11]. Handing over the fix leaves the applying to a person, and Akuity's On-Call Agent works the same way, pulling logs and recent changes to suggest a fix [9]. Akuity separately said on-call engineers are never paged for the Kubernetes application incidents the software catches and fixes on its own [20].
Hong Wang, Akuity's co-founder and chief executive, said agents have moved past writing code and into "how infrastructure changes get delivered and promoted to production" [13]. He added that the winners of the next phase will not be whoever adopts AI fastest, but whoever gives their agents a reliable path to do that work [14]. AI automations arrived on the platform last September [16], roughly 12 months before this launch [17].
A team weighing any of this has to ask whether the action reads or writes, and whether a named human's identity attaches at the moment it runs. Writes with no identity attached are the case Akuity is arguing against [3], and writes with an identity attached are what the policy layer is aimed at [6]. Reads with an identity attached are where the Major League Baseball result landed, and Goodness said, "This is already bringing value to the team, and we are looking to expand its adoption" [12].
What to watch
- Whether Akuity publishes pricing and availability for Agentic Control Plane and the MCP Server.
- Whether MLB moves past the triage work Goodness described to agent-initiated promotions.
- Whether other Kubernetes delivery vendors bind agent identity to a named human rather than a service account.