Agent protocols will consolidate the way network protocols did

The agent protocol landscape looks chaotic — MCP, A2A, SLIM, WebMCP, plus every vendor's proprietary variant — and the tempting response is to wait for a winner. I think that misreads how this movie ends. Protocol ecosystems under commercial pressure don't converge on a winner; they consolidate into a layered stack with narrow waists, the way networking did. The right bet is on layers and boundaries, not on any single protocol.

The layering is already visible if you squint: MCP handles tool access, A2A handles peer-to-peer task delegation, SLIM covers messaging transport, WebMCP claims the browser surface — four layers with reasonably clean boundaries, not four competitors for one job. More telling is who has shown up. The W3C runs a WebAgents Community Group; the IETF's OAuth working group and HTTP Message Signatures work are being extended for agents; the OpenID Foundation is mapping Token Exchange (RFC 8693) and agent-specific JWT claims onto delegation. When standards bodies, procurement departments, and identity architects arrive, protocol wars stop being decided by developer enthusiasm and start being decided by what composes — which historically ends in layering.

History also says which designs survive the consolidation. WS-* died of completeness: a committee stack that specified everything and composed nothing, losing to protocols that did one layer simply. SAML never escaped its assumption of a browser-mediated human session. OAuth's deepest assumption — a single principal per token — is exactly what multi-agent delegation breaks, which is why the current extension work is all about delegation chains: who is acting, on whose behalf, with what authority, across multiple hops. That multi-hop accountability gap — the cross-protocol trust gap — is the genuinely unsolved layer, and the financial-network analogies the research keeps reaching for are instructive: SWIFT turned O(N²) bilateral trust into O(N) network accreditation, and ISO 20022 spent twenty years migrating payments from free text to structured semantics. Expect the agent equivalent of both: federation brokers for trust, and a long grind toward shared semantics.

What does an architect do with this today? Treat protocol choice as a binding, not an architecture. Keep agent logic behind thin adapters — the A2A/MCP bridge patterns already demonstrate that a tool can be exposed as a task and vice versa, which tells you the abstractions underneath are convergent. And invest in the invariants that survive any consolidation: agent identity, delegation chains, audit trails, capability description. Whatever wire format wins its layer will need all four, so work spent there compounds while work spent on format-specific cleverness depreciates.

The concession is timing. ISO 20022's twenty-year migration is the base rate for semantic convergence, not the pessimistic case, and protocol fragmentation persisting for years is a live risk the convergence literature itself flags as an open problem. I could be early by half a decade. But notice that slow consolidation strengthens the layering bet rather than weakening it: the longer the wire formats churn, the more valuable protocol-independent identity, delegation, and audit become — they're the only investments that hold value in both futures. The teams that will regret this era are the ones welding their architecture to one protocol's shape in 2026; the ones that treat protocols like network engineers treated them — as replaceable layers under stable invariants — will barely notice the consolidation when it finally arrives, whichever decade that turns out to be.