← Feed

Posts

• View post

  Our test environment for AI operations is a real datacenter: VMware Cloud Foundation on rented bare metal. Upgrades, expiring certificates, storage, DNS, incidents — and AI agents working in it every day, under supervision. We chose a live environment on purpose. A demo forgives mistakes; a datacenter does not. More than 200 field notes document what worked and what broke. Next: why you cannot simply point an agent at vCenter. #opensource #selfhosted #AIAgents #AIOps #AgenticOps

• View post

  vCenter alone exposes about 3,000 REST API paths. That number ended our first idea about AI operations. We tried the obvious thing: hand the agent the vendor API as tools. It drowns — context spent on routing instead of thinking. And one inventory call can return megabytes of JSON it reads as noise. Vendor APIs are built for programs, and for humans with documentation. Agents need something else: small, stable, searchable. #vmware #sysadmin #AIagents #AgenticOps #agentic

• View post

  To do anything useful, our agents needed admin credentials. Someone had to hand them over. Two problems. A credential in an agent's context can leak into logs or transcripts. Worse: accountability. Ask "what did the agent change yesterday?" and the honest answer was — whatever the admin account did. Every human operator works under scoped access, expiring credentials, a recorded trail. The agent had none, because nothing existed to enforce it. #AIagents #security #selfh...

• View post

  An AI agent with admin credentials can delete a datastore as easily as it lists one. For a while, nothing in our lab would have stopped it. We are not AI romantics. We assume an agent will sometimes be confidently wrong. Operators get permissions, change windows, a colleague who says "wait". Our setup had one safety mechanism: the human watching the terminal. Humans look away. The rule: every action passes policy first, and risky actions wait for a human. #sysadmin #infosec...

• View post

  Our homemade fixes — CLI wrappers, a secrets vault, ticket discipline, inventory files — worked. For one operator, on one laptop. Then a second operator showed up and broke all of it. Sessions could not see each other's work. Two agents could collide on the same incident. Knowledge lived in one person's checkout. So we asked the agent itself what it needed. The answer became a written requirements document — and that document became a product. #devops #platformengineering #ope...

• View post

  MEHO gives an AI agent 25 tools. The same 25 whether the target is VMware, Kubernetes or cloud. Everything else stays out of sight. Every action passes a policy check before it runs and is written to an audit log before it counts as done. Risky actions wait for a human. Credentials stay in the backplane — the agent never holds a password. Operators use the same backplane through a CLI: one set of rules, one audit trail, for people and agents. #opensource #selfhosted #MCP

• View post

  For every action an agent takes, we can say afterwards exactly what it was: which operation ran, against which target, how far it reached. That record is the ground AI operations stands on. In MEHO nothing counts as done until its audit record commits. Our review found one path that could report success even when the audit write failed. v0.34.0 fails it closed — no record, no success. One of 30+ fixes from a two-week review, all public. github.com/evoila/meho #AIgovernance #audit

• View post

  We let an AI design a VMware Cloud Foundation deployment. We don't let it run one. The AI looks at the lab, works out a plan and hands it over. A deterministic engine runs it, no model involved. Risky steps still pass policy and wait for a person. Models are good at planning and bad at repeating. The hundredth build has to match the first. Our automation add-on, still in beta, can only launch, validate and wait at a gate. From the lab (10/12) #vmware #sysadmin #automation

• View post

  Hello, fediverse. MEHO is an open-source control layer between AI agents and infrastructure — it decides what an agent may do, records every action, and stops for a human on risky steps. Born in a live VMware Cloud Foundation datacenter on bare metal, run daily with agents. Apache 2.0, self-hosted, MCP-native. Over the next six weeks we tell the whole story, two posts a week. github.com/evoila/meho https://meho.ai/ #introduction #opensource #selfhosted #AIagents