Discovery Is Not Delegation

Agent infrastructure is becoming too large to wire together by hand. Tools, skills, agents, APIs, and context servers now live across teams and environments. Without a catalog, somebody has to know each resource exists, locate its endpoint, decide whether it is still current, and configure every client separately. That work does not scale, and it gets worse each time another team publishes another capability.

A shared discovery format attacks that coordination cost directly. Publishers can describe a resource once. Registries can index those descriptions. Agents and people can search across catalogs instead of relying on memory or bespoke connectors. Local owners can keep control of their own records while a common protocol makes those records understandable elsewhere.

That is useful infrastructure. It also creates a dangerous temptation: once an agent can discover a capability, teams start treating discovery as permission to use it. The lookup feels like the hard part because it used to consume human time. In production, the lookup is only the first question.

A catalog entry says what a resource is supposed to do and how to reach it. It does not prove the resource is healthy, that its description is accurate, that its owner still supports it, or that its behavior is safe for the job in front of the agent. Metadata can be approved and still become stale the next morning.

The difference matters because agents act on descriptions literally and at speed. A person who finds an unfamiliar tool may pause, inspect it, and ask around. A routing system can send work there thousands of times because the tags matched. Better search increases the speed of a good decision and a bad one alike.

Treat every record as an input to policy, never as the policy itself. The factory still needs current health, version compatibility, ownership, data boundaries, cost limits, and a precise account of what the caller may do. If those facts live only in prose attached to a record, you have made the catalog easier to search without making delegation safer.

The factory should turn a discovered capability into a bounded route. First it filters by policy: environment, data sensitivity, task type, budget, and required authority. Then it chooses among the eligible resources using evidence from real runs. Finally it issues the minimum permission for that job and records exactly what happened.

That sequence separates availability from suitability. A deployment tool may be visible across the organization and still be forbidden in production for this agent. A cheaper model may handle a routine classification while a harder decision earns a stronger one. A specialized agent may be the best semantic match and still lose the route because its recent failures crossed the acceptance threshold.

This is where the factory becomes more valuable than the model. The model can express intent and compare descriptions. The factory owns the rules that survive model changes: who can call what, when a route is allowed, what it may spend, and which evidence changes the next decision. Discovery feeds that system. It does not replace it.

Federated discovery makes a strong promise: resources can become findable across environments without moving ownership into one central platform. That promise holds only when the enforcement point remains with the owner. A remote catalog can tell an agent where a resource lives. The resource must still decide whether this caller, for this task, with this scope, gets through.

Revocation has to travel just as reliably as publication. If an owner withdraws a resource, changes its permissions, or marks a version unsafe, the factory cannot wait for a stale index to expire whenever it feels like it. The route needs freshness rules, short-lived authority, and a failure mode that closes the gate instead of improvising around it.

A boundary an agent can bypass by searching another registry is not a boundary. Federation expands the field of available machinery, which makes consistent policy more important, not less. The winning design lets discovery travel widely while authority stays narrow, explicit, and easy to withdraw.

Our prediction is straightforward: open discovery will help agents assemble more of their own working environment. People will spend less time maintaining lists of tools, copying configurations between clients, and directing routine work toward the right specialist. That coordination will become another repeatable factory function.

It will eliminate work and some of the roles built around that work. Human judgment moves higher: deciding which capabilities the organization accepts, setting the consequences a route may create, and owning the standard the output must meet. That is not the same as keeping a person in the middle of every lookup. A person supervising each selection is a manual router, not governance.

Build the catalog. Use the common protocol. Let agents discover what the organization has made available. Then put a harder machine behind the search result: policy before selection, narrow authority before action, independent proof before acceptance, and failure history fed back into the next route. Finding the worker is valuable. Delegating the work is the decision that runs the business.