In a multi-agent ecosystem, an agent needs to find other agents that can do what it needs. Discovery is the hardest unresolved A2A problem in 2026 — multiple proposed standards, fragmented adoption, real interoperability still scoped.
Static configuration
Each agent knows the URLs of the others ahead of time. Works for small, controlled deployments. Doesn't scale, doesn't adapt to runtime conditions. The starting point for almost every team.
Centralized registry
Single source of truth: agent X is at URL Y with capabilities Z. Like a service registry for microservices. Single point of failure; ownership question across orgs.
DNS-based discovery
SRV records advertise agent endpoints. Capability metadata in TXT records or pointer to manifest URL. Naturally federated, leverages existing internet infra. Underused but compelling for cross-org A2A.
Capability matching
Once you've found agents, pick the right one. Match required capability (free-text, schema-described, or both) against advertised. Score by past performance, cost, latency. Cache results.
Trust and verification
Discovery without trust is dangerous (man-in-the-middle attacks). Signed capability manifests; verified identity. SPIFFE/SPIRE for workload identity. Out-of-band trust establishment (you don't blindly trust a discovered agent).