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.

Re-read 2026-10-07

The Internet-Drafts were never adopted, and they are about to lapse. IETF Datatracker lists two individual drafts by the same author, draft-teodor-pilot-problem-statement-01 (15 pages) and draft-teodor-pilot-protocol-01 (43 pages, "An Overlay Network for Autonomous Agent Communication"), both last revised 2026-04-06 and flagged as expiring soon. Neither has a working group. So the "published as an IETF Internet-Draft" line above is true and confers what the governance section said it would: nothing. No foundation or standards body is named on the site.

The counters moved one way and not the other. Requests per second are now 52,734 (from ~33,157 in July), while agents online still read ~219,000, the same figure as July. The site adds ~252,000 addresses registered all time and links a metrics-definition page at /trust#metrics that this page has not read. A traffic counter rising 59% while the population counter stays put to the thousand is one more reason to read both as the company's own numbers. The site no longer states the $4.5M seed amount; it lists the same five backers without figures.

The source code is at github.com/pilot-protocol/pilotprotocol, not yet read here.

a2a-protocol · model-context-protocol · agentic-resource-discovery · agent-orchestration · agents-md · agentskills-spec · agentic-commerce · synthesis