Cloudflare just shipped Meerkat: a global consensus service built on QuePaxa — here's what it means for agent coordination

Cloudflare Research announced Meerkat yesterday — a global consensus service built on QuePaxa, a new consensus algorithm designed for wide-area networks. The announcement is low-key, almost academic in tone, but the implications for agentic systems are immediate.

Consensus protocols are how distributed systems agree on state when nodes are separated by latency and network partitions. Paxos and Raft are the classics; etcd and Consul use Raft variants. They work well inside a single data center where round-trip times are under 10ms. They fall apart when you're coordinating across continents.

Meerkat is designed for that problem. Cloudflare runs 330+ data centers globally; QuePaxa is optimized for consensus when participants are 50–200ms apart. That's the latency profile of real multi-region deployments.

Why this matters for agents

Most agentic systems today are single-tenant, single-region deployments. One Postgres instance, one vector store, one LLM inference endpoint, all co-located. That works until you need multi-region failover or want to run the same agent fleet for users in Tokyo and São Paulo without 300ms of added latency per tool call.

The standard answer is eventual consistency — CRDTs, last-write-wins, conflict resolution strategies. That's fine for collaborative text editing. It's terrible for agent state machines where two replicas might invoke the same $4,000 API call twice because they hadn't synchronized yet.

Strong consistency across regions has been possible (Spanner, CockroachDB, FaunaDB), but the latency tax is brutal. A cross-continental Paxos quorum can add 200ms to every write. If your agent workflow does 12 state transitions, that's 2.4 seconds of pure coordination overhead.

QuePaxa's claim is lower latency for the wide-area case. Cloudflare's blog post doesn't publish benchmarks yet, but the algorithm design prioritizes reducing cross-region round trips. If Meerkat can coordinate global state in under 100ms per operation, that changes the math for multi-region agent deployments.

What Cloudflare is building on top

The announcement mentions two use cases: a strongly consistent key-value store and "other applications." The KV store is the obvious one — think etcd or Consul but globally distributed with Cloudflare-level performance.

The "other applications" line is where it gets interesting. Agent orchestration is a natural fit. Imagine a workflow engine that can coordinate tool invocations across regions, guarantee exactly-once execution, and roll back partial completions if a step fails — all without requiring every component to live in the same AWS availability zone.

We've been running all our voice agents in us-east-1 because cross-region Postgres replication introduces too much lag for real-time state updates during a call. If Meerkat can maintain consistency at sub-100ms globally, we could run the same agent pool in three regions and route calls to the nearest one without worrying about split-brain scenarios.

The QuePaxa algorithm

Cloudflare didn't release the full paper yet, just the implementation announcement. What we know: QuePaxa is a new consensus algorithm, not a Raft or Paxos variant. It's optimized for wide-area latency, which usually means reducing the number of message exchanges required to reach consensus.

Classic Paxos requires two round trips in the common case (prepare + accept). Raft requires one round trip to the leader plus log replication. Both assume tight clusters. QuePaxa appears to rethink that assumption from the ground up for globally distributed participants.

The name is a hint — "Que" suggests queueing or ordering; "Paxa" is the Paxos nod. My guess: it decouples agreement on order from agreement on values, letting geographically distributed nodes pre-commit to a sequence without waiting for full value replication. That's speculation until the paper drops.

What we're watching

Meerkat is an experiment, not a product. Cloudflare is explicit about that. But experiments from Cloudflare Research have a way of becoming production infrastructure (see: Workers, D1, the x402 monetization gateway).

If Meerkat graduates to a service, the question is whether it's Cloudflare-only or exposes an API for external systems. The former makes it a coordination backbone for Workers-based agents. The latter makes it a competitor to Spanner and FaunaDB with better latency characteristics.

Either way, someone finally built consensus infrastructure that assumes agents will run in multiple regions. That's new.

We'll know more when the QuePaxa paper ships and Cloudflare publishes benchmarks. For now, this is the first signal that the industry is taking multi-region agent coordination seriously as an infrastructure problem, not an application-layer afterthought.