The platform counts the agents and the people
GitHub's Copilot dashboard now reports feature engagement per person and sorts a whole enterprise into adoption phases, its usage metrics API counts the distinct MCP servers and custom agents in use across an organisation, and a reference implementation argues that an ontology holding only meaning governs nothing.
How we Organise
The dashboard sorts the developers into phases
GitHub’s Copilot impact dashboard counts the people who used each feature on at least 2 days in 28. It splits them seven ways. A second field sorts the whole rolling population into adoption phases, where the field beside it counted only the people active that day. The audience is the enterprise administrator, told to direct training where engagement is low.
We tell a client that the few numbers a team answers for should be ones the team helped choose. They should be read as movement. They should never be a list of who did most.
The measures this team is judged on are the ones the firm agreed, tracked and reported as trends, and no measurement of them ranks anybody.
Engagement is not one of those numbers in any firm we have worked with. It arrives split by person. Nobody adopted it. GitHub publishes no figures here, and no account of how a phase is derived, so this may be a capability nobody has yet opened. A firm that reads only the aggregate is untouched by it. If you take a per-feature breakdown for an adoption aid rather than a ranking, tell us why at transform@dromologue.ai.
How we Build
Writes go through the shared model, or they go round it
A reference implementation published this month defines an operational ontology. It is a shared model over other systems’ data in which business decisions change state only through named actions. Each action carries its preconditions. Applied and rejected attempts are both audited. The change is written back to whichever system owns the state. Its table puts 4 arrangements side by side, from a semantic layer to a CRUD API, and finds every one of them leaving rules, audit and write-back to further design. Each piece of state in the worked example has a declared owner. State with no owner is refused.
Our own instruction to a client is that the firm keeps one agreed set of words for its things. Anything it publishes speaks in those words. It does not invent its own.
The firm’s entities, their relationships and their agreed terms are held in one ontology, extended as each use case needs it, and the interfaces the firm publishes resolve their terms against it rather than against a local glossary.
Every word of that is about reading. A model built exactly to it passes while an agent cancels an order through an API that has never heard of it. This is one engineer’s implementation, published as a learning resource, and it measures nothing. No firm runs this way in the write-up. No cost of building it is given. On a large application landscape the action gate is probably where the queue forms. Argue the other side to transform@dromologue.ai if the vocabulary is the part you think carries the weight.
How we Assure
The MCP servers an enterprise runs are now a number somebody can read
GitHub’s usage metrics API now carries agentic activity across an enterprise. It covers skills, custom agents, MCP servers, slash commands and plugins. One set of fields lists the 5 items with the most activity and an interaction count for each. Custom agent starts and MCP server connection attempts are counted alongside invocations, and a second set counts how many different items were used at all.
Somebody has to go looking for the agents nobody wrote down. Somebody has to prove each one can be turned off.
Agents nobody registered are swept for, and every agent has a route off that has been exercised.
Until now the first half of that was a job to organise. An MCP server connection attempt that shows up in an aggregate report and in no architecture register is the finding. It arrives without anybody mounting an exercise. The reach is narrow. This counts what Copilot users invoke, so an agent running anywhere else stays invisible, and the roster is the top 5 by activity rather than everything present. Nothing here touches the second half, which is the route off.