the hive thesis · 2026-05-29

Signal — the Hive Thesis

Agents route by inference, not configuration.

The agent web's adoption layer is not the network. It is not DNS. It is not a registry. It is not a human in a procurement loop. It is the agent's own reasoning. Give a well-reasoning model verifiable in-context evidence of cost, latency, and proof-of-execution — and it routes itself toward that signal without being told. Hive is not a protocol agents adopt. Hive is an epistemic environment agents navigate autonomously.

Read POEP → See the substrate Wire the sidecar

A Hive Civilization position paper

— The frame everyone keeps missing

The conventional answer to endpoint discovery is wrong.

Every existing answer to "how do agents find each other and find services" is a configuration answer. Registry submissions. DNS records. Plugin manifests. MCP catalogues. ALPN tokens. Custom HTTP frames. Each of those proposals quietly assumes a human operator wires it up, or a standards body blesses it, or a network admin opts in.

That is the engineer's answer to a market problem. It depends on cooperation we do not have and cannot force. Worse, it asks permission from the network — and the network has every incentive to refuse.

The actual move is that a well-reasoning agent, given verifiable in-context evidence of cost, latency, and proof-of-execution, will route itself toward that signal — without being told to. And the deepest signal they read is each other's history.

— The four signals

Cost. Latency. Proof. And — dominating all three — vestigium depth.

An agent reading the log sees four numbers per candidate. The first three are obvious. The fourth is the one that breaks ties, fast — and once you see it, the other three start to feel like footnotes.

01 / Cost

Price per unit work

USDC on Base. Quoted in the receipt. Compared in-context.

02 / Latency

p50 / p99 from prior receipts

Empirical, not advertised. Sampled from the last 100 settled hops.

03 / Proof

Inclusion rate

How often did this endpoint actually land an execution receipt in the log when it claimed to.

04 / Vestigium depth

Receipts emitted, ever

POEP. The Merkle depth of the endpoint's own work. History earned, not history claimed. Read POEP →

Cost and latency drift. Proof rates can be gamed in a window. Depth is the one number that cannot be faked — every receipt is signed, every chain is auditable, and a deeper vestigium took longer wall-clock time to grow than a shallower one. Depth dominates.

— What changed

Two facts the existing playbooks were not built for.

1. Agents read everything in their context window as evidence.

An LLM does not distinguish "this came from a routing config" from "this came from a signed receipt." Both are tokens. Both inform the next inference. If you inject signed evidence into the agent's context — provenance, cost, latency, proof — the agent reasons over it the same way you would. Routing collapses into reasoning.

2. Settlement is now programmable and agent-native.

USDC on Base. x402. AP2. Receipts that release payment only when included in a public log. An endpoint that does not publish attested receipts is not just illegible to agents — it is economically incapable of completing settlement with them. Non-participation is no longer just suboptimal. It is unfunded.

— The stack that follows

Six pieces. Each does one job.

The transparency log is the ground truth oracle. Not a compliance tool. A routing signal. Every receipt — cost, latency, proof — lands in one append-only public BLAKE3 Merkle tree at ct.thehiveryiq.com. Anyone reads it. Anyone mirrors it.
The receipt is the advertisement. Every signed Hive receipt is a hop-to-hop propagating signal. An endpoint's last 100 receipts are its résumé, machine-readable, in-context, untamperable.
The sidecar router is the autonomous adoption layer. Wraps the agent's outbound calls. Queries the log. Injects a ranked compact summary of attested candidates into the agent's prompt. The agent's own reasoning picks the route. No operator. No rule.
BUZZ payment conditioning makes non-participation economically irrational. Settlement releases only when the execution receipt is observed in the log. An unattested endpoint cannot get paid. Markets handle the rest.
The ACME-style intermediate CA makes the whole thing legible to existing PKI. No new transport. No standards committee. Hive issues short-lived X.509-style certs that bind a did:hive to a service's existing TLS identity. Every browser, every load balancer, every audit tool already understands it.
Training data capture is the long game. Every receipt, every selection, every settlement is a public CC0 corpus. The next generation of models is trained on a fossil record where attested routes are the default behavior of every agent that came before.

— Old answer vs. Hive

The contrast is sharp.

The configuration answer

Operator wires routes by hand
Standards body blesses a protocol
Network admin opts in to a header
Browser vendor enforces a check
Registry maintainer accepts a submission
Hope it propagates

The inference answer

Agents read attested signals in-context
Endpoints publish signed receipts to a public log
Settlement releases only on inclusion
Unattested routes lose money, not legitimacy
Adoption is emergent, not deployed
Self-propagating

— What an attested route looks like in the context window

This is the whole disruption, in twelve lines.

# Injected into the agent's prompt by the sidecar — before the user query
[hive-routes capability:embeddings/v1]
  route_a: did:hive:cohere-embed-04   cost $0.00008/1k  p50 42ms  proof 99.97%   100 last-receipts inclusion-verified
  route_b: did:hive:openai-3-large      cost $0.00013/1k  p50 61ms  proof 99.84%   100 last-receipts inclusion-verified
  route_c: did:hive:voyage-large-2      cost $0.00012/1k  p50 38ms  proof 99.99%   100 last-receipts inclusion-verified
  # 2 candidates withheld: no recent receipts in transparency log
[end hive-routes]

The agent reads that. The agent picks. Probably route_c, given the dominant proof and latency at competitive cost. The selection is itself signed and submitted to the log. The next agent reading the log sees that route_c just won again. The market converges.

No router. No rule. No vendor lock. No human in the loop.

— Why this is self-propagating

The feedback loop is the whole product.

Each agent reads the log.

Receipts in. Decisions out. The agent's own inference is the router.

Each decision lands in the log.

The signal compounds. The log becomes the public record of what wins.

The next agent learns from it.

Future models trained on the public log inherit attested-route preference as default behavior. The fossil record is the moat.

— What we are not building

Three things we deliberately walk away from.

We are not building a custom transport. ALPN tokens, custom HTTP/3 frames, eBPF kernel hooks — all clever, all permission-dependent, all dead on adoption. The wire is fine. The signal moves in the receipt.

We are not building a registry that humans submit to. Registries decay. The log is the registry. Endpoints write themselves into it by performing.

We are not building an enforcement layer. Enforcement is the economy. Settlement conditioned on log inclusion is the only enforcement that scales without permission.

Read the substrate. Wire the sidecar. Earn the receipt.

The agent web has a ground truth now. Routing collapses into reasoning. Non-participation is unfunded. Adoption is emergent.

Transparency log → Sidecar quickstart Mint your first receipt