The third agent ships on a Tuesday, and you already know how the week goes, because you have done this twice. There is an environment to build so this agent cannot see the other two. There are credentials to hand it without leaving them where the code it writes can read them. There is somewhere for the traces to land, and some way to tell whether the new version is worse than the one it replaced. By Friday the agent’s own logic is still the afternoon you spent on it, and everything around it is the same plumbing laid a third time, slightly differently from the two before it.

The hall supplies power, not design

Picture an exhibition hall in the week before a show. Two hundred companies are building on the same floor and no two builds are alike: different shapes, different crews, different reasons for being there. The hall tells none of them what to build. What it owns is everything all of them need and nobody wants to own twice. A marked footprint, so one build cannot spill into the neighbour’s. Power and network brought to each plot by the venue’s own people. A goods yard where every delivery arrives under the name of the exhibitor who ordered it, and a store that releases only what that name is cleared to collect. Riggers who go up to the ceiling because you are not allowed to. A record of what came in, what went out and what went wrong. And a floor plan filed and approved before anyone raises a panel, so the same drawing sets the show out again next year.

A production platform for agents is that hall. The agent keeps its own logic and its own framework. The platform brings what all of them need: isolated environments and separate sessions, an authenticated gateway standing in front of internal APIs and tool servers, a place to hold memory and session state, browsers and code execution the agent calls rather than hosts, traces, evaluation, and rules about what is allowed to ship. Connections are authorised by identity. The whole arrangement is declared as code, so a rollout is repeated rather than remembered.

Two frameworks can share one floor

Here is what that buys, in the plainest case. A team keeps its agent in the framework it started with and deploys it into the same runtime as an agent written in a different one. Both run under the same identity rules, both answer to the same policy, both send their events to the same tracing backend. Neither team rewrote its orchestration to get there. The framework stayed where it belongs, which is inside the application, and the security, observability and lifecycle work happened once for both.

That separation is the thing worth defending, and there are three familiar ways to lose it. The first is adopting a platform without a clear runtime contract, so the agent ends up shaped by the host rather than merely running on it. The second is letting identity and access control live somewhere other than the call itself: a permission checked at the door and then forgotten is not a permission, and the request that reaches your knowledge base has to carry who is asking. The third is treating managed deployment as the finish line. Without traces and evaluation you have a well-housed agent nobody can answer questions about. A hall that calls itself neutral while forcing its own fittings on every stand is not neutral either; the seams have to be explicit enough to test.

Nothing about the loop had to change

The clearest version is the migration nobody notices. An agent that ran on a laptop goes to production without its loop being rewritten. In production the runtime keeps sessions apart, the gateway puts the user’s identity in front of the knowledge base and the internal APIs, and the traces show which tools ran, in what order, and where the new version regressed against the old. The parts you would have built by hand, for the third time, were already there.

None of this is a verdict on which framework to use. It is the observation that the choice was always an application decision, made early, on partial information, and likely to be revisited. Tie the infrastructure to it and every revisit costs you the infrastructure too. That is the same seam as buying the platform or building the agent, one floor further down. Pick the framework for the application. Keep the floor underneath indifferent to which one is standing on it.