Here's a pattern I've watched happen at nearly every enterprise running more than a handful of AI projects. Team A builds a Salesforce read-only connector for their agent. Three months later, Team B builds the same connector. Team C is halfway through building it when someone mentions at an all-hands that two other versions already exist. Nobody uses either of the first two because nobody knew they existed, and by now Team C's version has different behavior anyway.
This is not a technology problem. It's a sharing problem. And the fix is the producer/consumer flywheel.
The flywheel#
The thesis is simple. Your existing engineers, PMs, security folks, and operators are building agentic capabilities every day. They're just doing it in isolation, often without sharing what works. The flywheel formalizes the sharing without centralizing the building.
Producers publish assets to a shared internal marketplace. Consumers discover and use those assets instead of building from scratch. When a consuming team ships faster because of an asset a producing team published, the producing team gets credit in the next planning cycle. That credit is the economic engine most programs forget to build.
Six asset classes flow through the marketplace: agents (purpose-built, with declared scope and owner), workflows (multi-step orchestration), MCP tools (scoped wrappers around enterprise systems), policy templates (reusable governance patterns), eval harnesses (test suites for agent classes), and audit queries (saved analysis patterns for SecOps and compliance). All six are things your teams are already building. The marketplace just makes them discoverable and reusable.
The trap inside the reuse metric#
Gartner has a calibration worth internalizing: "Reuse is a good outcome, but a bad goal." If you optimize for asset count, you've built a vanity metrics machine. The RPA CoE wave died on this rock. Teams published hundreds of assets nobody consumed because the incentive was publication, not adoption.
The right question at every quarterly review is not "how many assets did we publish?" It's "did our consuming teams ship faster because of those assets?" That's the only number that tells you the flywheel is actually spinning.
The deprecation ritual#
The other thing most programs skip is the sunset discipline. Without it, your marketplace becomes a graveyard within eighteen months. Assets with no consumption in 90 days get flagged. Assets with no consumption in 180 days get sunset. The Asset Curator role, which is an assigned responsibility on your existing team, runs this review weekly.
This is not bureaucracy. It's the thing that keeps the marketplace credible. A marketplace where half the assets are stale is a marketplace people stop trusting. And a marketplace people stop trusting is a marketplace nobody uses.
What federation actually means#
The flywheel works because the central methodology owns the substrate and the federated practitioners own the use cases. The moment central also owns the use cases — "we'll build the agents centrally for consistency" — you've rebuilt the 2018 RPA CoE. Same org chart, different vocabulary.
Federation means Team A publishes what they built and Team B consumes it. It means the methodology team measures cross-team consumption and rewards the producers. It means the marketplace has discoverability and the deprecation ritual has teeth. What it does not mean is a central shop that builds things and hands them to the business.
The difference sounds subtle. In practice it's the difference between a flywheel that compounds and a bottleneck that doesn't.
Part five of a series based on the Agentic Adaptation Playbook. Previously: the most important AI meeting you'll have this quarter. Next: why security is the engine of your agentic migration, not the brake on it.
