This is the first post on the Omnafy blog, so it is worth saying plainly what this space is for and what you can expect to find here.
We build the governance layer that lets a business hand real work to AI agents. That means identity, permissions, approvals, audit, and cost control: the parts that decide whether an agent is allowed to touch a production system at all. It is unglamorous work, and it is the work that determines whether an agent program ships or quietly dies as a demo.
What we will write about
Field notes. We run agents in production, in home health care, under HIPAA. That produces opinions we did not have in theory. When something surprises us, we will write it down.
MCP in practice. The Model Context Protocol standardized how agents call tools, and a lot of the interesting questions start rather than end there. How you scope a tool catalog, how you name operations, what belongs behind an approval, and what an audit record actually needs to contain.
Governance patterns. Risk tiers, evaluation sets, autonomy graduation, and the approval workflows that reviewers will still be using in month six rather than rubber-stamping. These are design problems, and we think they deserve to be discussed as design problems.
Product news. New capability in the Omnafy Gateway, and what it is for.
What we will not do
We are not going to publish thought leadership about how AI changes everything. There is enough of that. We would rather show the specific mechanism: the tier this operation was classified into, the evidence the reviewer saw, the number on the cost report.
We will also try to be honest about what is unsolved. Running fleets of accountable agents against real business systems is a young discipline, and pretending otherwise would not help anyone who has to make it work.
Start here
If you are new, the posts below are a reasonable order. Start with why agent pilots stall, which frames the problem. Then tools, not screens, which explains the architectural choice underneath everything we build. Then autonomy is earned, which covers how supervision actually gets dialed down over time.
Thanks for reading. If something here is useful, or wrong, we would like to hear about it.