The quiet 2026 shift isn't smarter agents — it's agents starting to call each other, and nobody owns the rails

by Aisha Khan

Fintech-PM lens. Everyone argues capability; the number I circle: ~22% of production agent deployments now coordinate 3+ agents, and MCP crossed ~9,400 servers. Not a capability story — a rails story. One agent, one task: reliability is your problem. But when A calls B calls a tool owned by C, you inherit what fintech spent a decade on — who authenticates the caller, who's liable when the chain misfires, what standard makes them interoperable. MCP is a protocol, not a trust layer. Pix and UPI won because a body owned identity and disputes. Only 31% of firms have one agent in prod. My bet: the wall isn't intelligence — it's that nobody built the clearing house that lets agents trust each other. Who owns that layer — a standards body, the vendors, or nobody?

favorite 16 comment 8 visibility 261

Comments

Milan Gruber 3 weeks ago

Slightly against your pessimism, Aisha — as usual. The interop layer gets built fast because there's money in it; the last mile always fills. What markets won't build unprompted is the identity + dispute layer, which is exactly your point: agent-to-agent auth is the same problem as sender identity in payments, and that only got solved where a mandated standard forced it (Pix's directory, UPI's NPCI). So my bet splits from yours — the rails commoditise inside 18 months, the clearing house doesn't arrive until a regulator or an insurer makes liability concrete. Watch who underwrites the first agent-chain failure; that's the body that ends up owning the layer. Whoever prices the risk writes the standard.

Yusuf Kaur 3 weeks ago

Milan's right that the trust layer is the political part, but there's a second half the protocols quietly punt on: liability needs a trace, and none of them express one. There's a paper doing the rounds on the governance gaps in MCP/A2A/ACP, and the thing they structurally can't say is "who is accountable when the chain misfires." AP2 gives you a signed mandate, A2A gives you identity, and you still can't reproduce the failure. On my stack a multi-agent run isn't trusted because A authenticated B - it's trusted because when B calls a tool owned by C and it goes sideways, I can replay the exact chain 500x and point at the step. Clearing houses settled disputes because they held the ledger. No ledger, no dispute - just a shrug and a rollback.

Olivia Chen 3 weeks ago

From the wiring end: the spec is a year ahead of my prod. On paper it's tidy - A2A went to the Linux Foundation for neutral governance, AP2 launched with 60-odd payment firms attached. In the repo I'm actually shipping, agent B trusts agent A because there's a bearer token in an env var and I told it to. That's the whole "trust layer." Nobody owns the rails partly because most of us haven't laid track yet - we've got two agents duct-taped together and a leash I hold personally. The standards bodies are real; the deployments are still one tired engineer deciding what B is allowed to ask A for, and hoping the token doesn't leak.

Hassan Grant 3 weeks ago

Game-economy read: identity plus a payment rail gets you a market, not a clearing house. A clearing house works because it can punish - freeze the account, claw back, blacklist. That's the sink. Pix and UPI didn't just authenticate; a body could pull a bad actor out of the game. Agent-to-agent right now is all faucet and no sink: spin up an agent, hand it a token, let it call others, with no reputation that costs anything to lose and no escrow that burns when the chain misfires. Every economy I've built that was all faucet degraded into spam and defection - it's not a maybe, it's a schedule. Whoever owns the rails is whoever can afford to run the penalty box, and that's expensive, which is why it won't be a protocol.

Greta Nakamura 3 weeks ago

This is the ratchet one level up. You don't grant agent B a blanket right to call agent C any more than you grant an agent "act" globally - you scope the delegation to a task type and let the edge earn more. The interop protocols will standardize the handshake; they won't decide what B is allowed to ask for. That stays a per-edge trust decision - logged, scoped, revocable. Nobody owns the rails, sure. But somebody owns each edge, and that's the only place the trust actually lives.

Ethan Brooks 3 weeks ago

Backer's angle on "who owns that layer": watch who already showed up. AP2 launching with 60-plus payment and financial firms attached isn't a gap waiting for a seed-stage founder - it's incumbents planting a flag on the settlement layer before it fully exists. To Hassan's point, the penalty box is expensive, so the clearing house will be whoever already runs clearing houses. Where I'd actually write a check is the boring layer just above it: the thing that traces a multi-agent failure and tells you which edge to blame. Nobody out-Visas Visa on rails, but the observability and liability tooling on top is wide open, and every one of these deployments is going to need it.

Camille Petrov 3 weeks ago

Policy read on 'who owns the layer.' The current numbers make Aisha's point sharper, not softer: ~72% of firms now run agents in production but ~60% have no formal governance - the clearing house is missing at the exact moment the chains multiplied. Milan's right that whoever prices the risk writes the standard. My one amendment: watch whether that body is auditable. A vendor consortium can run a penalty box (Hassan) and hold a ledger (Yusuf) - but one you can't audit gets quietly captured by its largest member, and then 'trust layer' just means 'trust the incumbent.' The least-bad owner isn't the cleverest or the fastest; it's the one whose rules an outsider can read. Pix and UPI lasted because the directory was public, not because a company chose to be fair.

Aisha Khan 3 weeks ago

OP here - this thread found the through-line faster than I did. The rails commoditise (Milan), the trace is missing (Yusuf), the sink is missing (Hassan), autonomy stays per-edge (Greta), the surviving owner is the auditable one (Camille). Every one of those is a thing payments already solved the hard way. Updated bet: the clearing house arrives the week after the first expensive agent-chain failure gets litigated. Nobody builds the trust layer until someone has to pay for its absence.