Abstract
<title>Abstract</title> <p>API integration burdens every consumer with provider-specific endpoints, schemas, versions, credentials, and error conventions, reworked at every provider change. We present SA2A (System Agent-to-Agent), an architecture in which one System Agent at each system's boundary accepts signed, intent-level envelopes, discovers capabilities dynamically, enforces deterministic local policy, and returns signed, field-filtered responses. Its defining property is mutual zero-trust: because any peer agent may be compromised, responses pass an entry gate at the requester exactly as requests do at the provider. A three-agent prototype (473 automated tests) was evaluated against a standard-toolkit REST baseline. All eighteen designed attacks failed, including six a request-only gate would admit: forged, replayed, over-broad, container-wrapped, and instruction-carrying responses. Adding a capability cost zero client-side steps against six for REST; under a breaking change an unmodified REST client broke on all five operations, one silently, while the SA2A client failed loudly and recovered without code edits. The price is + 15.6 ms per call and + 624 bytes per message; decomposition attributes 84.6% to durable audit and nonce writes, 1.6% to cryptography, and 0.1% to policy evaluation. Two ablations, over six provider layers and six response-gate checks, find one provider layer redundant and every gate check load-bearing.</p>