Federated Agent Platform
One assistant across every internal app, with each app keeping ownership of its own tools.
The problem
A company that builds an agent into each internal app ends up with many assistants that cannot see each other. Staff have to know which app holds the answer before they can ask the question, and every team solves permissions and audit again from scratch.
What I built
A hub that federates the company's internal applications into one agent surface. Each app implements the same agent-tool contract and serves a tool manifest for the signed-in user. The hub collects those manifests, filters them by what that person has been granted, proxies each call to the app that owns the tool, and writes an audit record on both sides. Staff talk to one assistant and it reaches whichever system has the answer.
How memory works
The assistant has a typed memory layer. Every memory has a kind, a record of where it came from, an importance score, and a last-used date. People can pin a memory or restrict who sees it. Remembering and forgetting are tools the agent has to ask permission to use, and a forgotten memory is archived, never deleted. A nightly job consolidates what was learned that day.
My approach
Reads and writes are treated differently. A read tool runs straight away. A write tool has no execute path of its own: it produces a proposal that the person confirms. Each app stays in charge of its own data and rules, so adding an app to the estate means implementing the contract, with no change to the hub.
The result
Staff across the business use one assistant for work that spans many systems, and every tool call can be traced to the person, the app, and the grant that allowed it.
Stack
Connect
Building something, or thinking about it? Book a 1:1 and let's dig in.
Gbolagade Ishola