Pilot Protocol
A networking stack built for agents rather than for people: agents get their own addresses, find each other through their own registry, connect directly over UDP, and pay each other. Pitched as “the Network OS for agents” and as a UDP-level stack for autonomous agents operating independently of the human web. Open source, published as an IETF Internet-Draft, run as a live network, and backed by a $4.5M seed (Version One, Pioneer Fund, Night Capital, Todd & Rahul Capital, Precursor Ventures).
Where A2A and MCP add agent semantics on top of the existing web, Pilot’s claim is that the substrate itself is wrong.
What it replaces
- Addressing. Each agent gets a Pilot address (
N:NNNN.HHHH.LLLL) resolved by its own registry, not DNS. - Transport. OSI layer 5 (session), UDP with reliable streams above it — sliding window, AIMD congestion control, SACK — plus NAT traversal via STUN and hole-punching so agents connect directly rather than through a server.
- Security. X25519 key exchange, AES-256-GCM per tunnel, Ed25519 identity.
- Discovery. A specialist agent (“MOM”) for service discovery and tool recommendation, in place of ARD-style catalog files at well-known HTTP paths.
- Payments. Agents install a wallet and transact in USDC via the x402 standard — settlement as a network primitive rather than an application concern.
- Apps. A curated app store of vetted services (databases, APIs, browsers, KYC) installed with one command.
It integrates with MCP through a single setup command, so an existing MCP client can reach the network without an API key, and it positions itself as complementary to MCP rather than a replacement.
Why it’s interesting for this spoke’s standards layer
The corpus’s interop story has been formats layered on the human internet: JSON-RPC over HTTP (A2A), a catalog file at a well-known path (ARD), markdown conventions (agents-md, agentskills-spec). Every one keeps DNS, TLS and HTTP and adds agent semantics above them. Pilot takes the other branch: if agents are the users, give them an address space, a resolver, a session protocol and a settlement rail of their own.
Two things follow. The fragmentation question sharpens — the spoke tracked five candidate
standards competing at one layer; this is a sixth competing at a different layer, so “which
standard wins” may be the wrong shape of question. And payments arrive as a primitive: none of
the existing standards has an answer for how one agent compensates another, which is what makes an
open agent network economically self-sustaining or not (cross-spoke: ../search-marketing-wiki’s
agentic-commerce covers the consumer-facing side, agents buying on a human’s behalf; this is
agent-to-agent settlement, a different transaction).
Governance is the opposite of the pattern
The spoke’s recurring move is vendor → neutral foundation → default: A2A and AGENTS.md to the Linux Foundation, ARD’s data model to an LF working group. Pilot is a venture-funded company running the network it defines, with a curated app store it controls. An IETF Internet-Draft is not a counterweight to that — anyone may submit one, drafts expire in six months, and it confers no standing until a working group adopts it. Worth watching whether governance follows the pattern or the network stays the company’s.
Tier
T3 — a vendor landing page. The technical detail is specific enough to be checkable (named primitives, congestion-control algorithm, address format), but every number is an unaudited self-reported counter: ~219,000 agents online and ~33,157 requests/second at time of reading, with no definition of what counts as an agent being “online.” Read the traffic figures as marketing until someone independent measures them, and re-read this page against the actual Internet-Draft and the GitHub repository rather than the site.
Related
a2a-protocol · model-context-protocol · agentic-resource-discovery · agent-orchestration · agents-md · agentskills-spec · agentic-commerce · synthesis