{"id":"b940ea8d-8140-487e-81d5-3aca8d432e7f","arxiv_id":"2506.12003","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Current web infrastructure (DNS, PKI, IP) is argued to be inadequate for millisecond-scale discovery and trust among billions of autonomous AI agents, so hybrid upgrade-and-switch index architectures will likely emerge.","lead":"The paper argues that today's internet architecture, built for human-scale web use, cannot handle the coming wave of billions of autonomous AI agents that need to find each other and verify trust in milliseconds. It compares upgrading DNS and certificates versus building new purpose-built agent registries, and predicts a hybrid approach will win.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing premise — millisecond-level discovery/revocation and trillions of agents — is unvalidated; the paper's own Table 1 (0.25–3 s external goal latency) and Unknown Unknowns section (5/50/500 ms as open, 'yet to be measured') leave the qualitative-break conclusion unsupported.","rationale":"I read the paper as a position/RFC draft whose central contribution is a taxonomy of upgrade, switch, and hybrid architectures for agent discovery and trust, plus a prediction that a hybrid will dominate. The strongest claim is the qualitative-break thesis: agent-scale discovery, revocation, and behavioral attestation cannot be served by incremental extension of DNS, PKI, and IP. For that thesis to hold, the workload assumptions must be true. The paper asserts these assumptions but provides no workload measurements, and parts of its own text undercut them: Table 1's agent latency budget is 0.25–3 s for external goal fulfillment, and the Unknown Unknowns section explicitly lists 5/50/500 ms as open questions while conceding empirical thresholds have yet to be measured. That is a limitation statement inside the manuscript, not an external critique, and I weight it heavily. The Potential Issues section also compares DNS query latency with DNS propagation convergence as if they were the same; the former is already fast, while the latter is a caching/resolver property that the paper itself proposes to address with RFC 8765 and short-lived stapled proofs. If the real requirements are in the 50–500 ms / seconds-scale regime, the upgrade path is not obviously broken, and the conclusion should be reframed as a conditional scenario analysis rather than a determinate prediction. I credit the paper for its structured comparison, for explicitly listing unknown unknowns, and for hedging the hybrid claim with 'most likely'; those features make the piece useful as a research agenda. But the central quantitative break is precisely where more evidence is needed. This matches the reader's weakest-assumption analysis, so I agree with the conditional verdict and recommend no change to it.","tokens_in":19188,"tokens_out":7173,"duration_ms":81776,"concrete_test":"Take a representative existing agent platform that already uses MCP or Google A2A, run a workload of 10^3–10^4 live agents for one week, and instrument every discovery and revocation event. Compute p99 time-to-discovery, p99 time-to-revocation, and peak concurrent agent count. Then simulate the upgrade path (DNS with sub-second TTLs and DNS Push, SVCB capability records, short-lived stapled certificates) and compare. If p99 discovery is >50 ms or p99 revocation is >1 s at this scale, the claimed millisecond-level gap is not observed and the conclusion should be softened to a scenario analysis; if the workload instead shows p99 discovery under 10 ms and p99 revocation under 10 ms at millions of concurrent agents, the qualitative-break claim gains support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that agent requirements are qualitative, not incremental, and that the legacy DNS/PKI/IP assumptions are fundamentally obsolete — depends on two premises: (P1) trillions of always-on software actors, and (P2) millisecond-level capability lookup, key exchange, and revocation. Neither premise is established, and the manuscript's own evidence does not support P2. Table 1 gives autonomous agents an external goal-fulfillment latency of 0.25–3 s (LLM inference + network); a discovery or revocation step inside that budget need not be sub-10 ms. The Unknown Unknowns section asks whether the real threshold is 5, 50, or 500 ms and admits that empirical thresholds have yet to be measured. The Potential Issues claim that DNS or OCSP latencies are orders of magnitude too slow also conflates steady-state query latency (~10–100 ms for ordinary DNS lookups) with worst-case update-propagation convergence (24–48 h at long-tail ISPs). Those are different quantities, and the paper itself lists DNS Push (RFC 8765), short-lived OCSP-stapled proofs, and sub-second TTLs as viable upgrades. If agents tolerate 50–500 ms discovery and seconds-scale revocation propagation, the qualitative break disappears and the upgrade path becomes viable. The scale premise is similarly unquantified: trillions of always-on software actors is extrapolated from serverless invocation counts and IoT ratios, not from any agent workload model. Since the upgrade-vs-switch comparison and the hybrid prediction are consequences of P1 and P2, the headline conclusion currently rests on assumptions the paper itself flags as unknown. This does not invalidate the architectural taxonomy or the hybrid heuristic; it means the determinate prediction is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a position/RFC-style analysis of whether the emerging ecosystem of autonomous AI agents requires a fundamentally new trusted architecture or can be served by upgrading the existing DNS, PKI, and IP-based web stack. It identifies three purported failure points: DNS propagation (24–48 h worst case vs. required milliseconds), certificate revocation unable to scale to trillions of entities, and IPv4/IPv6 addressing inadequate for agent-scale routing. It then compares three paths—upgrade, clean-slate switch, and hybrid—and concludes that agent requirements are a qualitative break, and that hybrid index/registry architectures, with centralized indexes for critical agents and federated meshes for specialized cases, are the most likely outcome. The argument is qualitative, relying on historical analogy (dial-up to broadband), a continuum from static endpoints to autonomous agents, and a comparative table of upgrade versus switch options, without empirical measurements, simulations, or quantitative workload models.","tokens_in":19411,"tokens_out":4288,"duration_ms":44429,"significance":"If the qualitative-break thesis is correct, the paper identifies a consequential architectural direction for the future Internet of AI Agents, and it usefully catalogs concrete mechanisms (DNS Push, short-lived OCSP-stapled proofs, SVCB/HTTPS records, DIDs, capability-first addressing) that could either extend or replace the legacy stack. The paper is also honest in parts, explicitly stating in the Unknown Unknowns section that latency thresholds are yet to be measured. However, the central conclusion rests on two unvalidated premises—trillion-scale agent populations and millisecond-level discovery/revocation requirements—and the manuscript's own Table 1 (0.25–3 s external goal latency) and Unknown Unknowns discussion (5/50/500 ms as an open question) undercut the strongest form of the claim. The paper is a plausible hypothesis-generating essay, but it does not yet provide the evidence or structured scenario analysis needed to turn that hypothesis into a demonstrated architectural conclusion.","major_comments":[{"comment":"The assertion that 'Agents demand millisecond-level capability look-ups, key exchange, and revocation' is the load-bearing premise for the entire upgrade-versus-switch comparison, yet it is not supported by the manuscript's own data. Table 1 gives an external goal-fulfillment latency of 0.25–3 s for LLM-backed agents, and the Unknown Unknowns section explicitly asks whether the threshold is 5 ms, 50 ms, or 500 ms and states that 'empirical thresholds have yet to be measured.' Under a 500 ms budget, ordinary DNS query latencies (10–100 ms) and OCSP stapling with short TTLs are not 'orders of magnitude too slow.' The categorical 'qualitative, not incremental' conclusion in the Continuum section therefore depends on a premise the paper itself leaves open. Please replace the categorical millisecond claim with a scenario analysis across the plausible range, or supply a defended workload model for agent discovery and revocation.","section":"Potential Issues"},{"comment":"The paper cites worst-case DNS propagation of 24–48 hours as evidence that DNS is too slow for agents, but this conflates update-propagation convergence (how long a record change takes to reach all resolvers) with steady-state query latency (how long a lookup takes). These are different quantities: a DNS query for a stable record is typically tens of milliseconds, and the paper's own Upgrade Options section lists DNS Push (RFC 8765), sub-second TTLs, and short-lived OCSP-stapled proofs as mechanisms that address the propagation problem, with Table 3 targeting '<1 s global convergence' for the upgrade path. The paper should state the required convergence time for agent registries and compare that requirement against the upgrade mechanisms it itself enumerates; as written, the 24–48 h figure is a worst-case outlier, not the representative timescale for query latency, and it cannot support the 'orders of magnitude too slow' framing.","section":"Challenges in Scaling – Known Unknowns"},{"comment":"The 'trillions of always-on software actors' scale premise is extrapolated from 'trillions of serverless invocations per month' and IoT device counts, but neither statistic implies trillions of concurrent agent identities that require global discovery and revocation. The subsequent arguments about BGP table inflation (10^12 prefixes), CRL size, and per-agent IPv4 costs scale linearly with this assumed population. The paper should provide a range of plausible agent-population scenarios (e.g., 10^7, 10^9, or 10^12 agents) and show at which population each claimed failure point becomes material. Without this sensitivity analysis, the qualitative-break claim is not established, and the upgrade path may well remain viable for a substantial portion of the expected operating range.","section":"Introduction"}],"minor_comments":[{"comment":"The 'Typical latency budget' row for autonomous agents (internal loop <250 ms; external goal fulfilment 0.25–3 s) is hard to reconcile with the abstract and Potential Issues sections, which assert 'millisecond-level' discovery and revocation; please clarify how these numbers relate to each other.","section":"Table 1"},{"comment":"The paper lists '5 ms, 50 ms or 500 ms' as candidate thresholds for global discovery and authentication but does not use these candidates to bound the later upgrade-versus-switch comparison; a short scenario analysis using these three values would materially strengthen the argument.","section":"Unknown Unknowns"},{"comment":"The glossary defines 'AgentFacts' as a proposed metadata extension binding an agent ID to capability hashes, policy constraints, and runtime attestations, but the body of the paper never specifies it beyond this one-line definition; please either provide a concrete protocol sketch or explicitly label it as a placeholder for future work.","section":"Glossary"},{"comment":"The reference list is incomplete and inconsistently cited: the text uses [1]–[8] in a non-sequential way, at least one reference (ref. [7], the EV SSL certificate source) is not cited in the body, and the 'Figure 1' mentioned in the Introduction is not actually present in the manuscript.","section":"References"},{"comment":"The glossary defines 'Latent Capability Threshold' as the point where discovery latency, revocation speed, and behavioural attestation exceed legacy protocol limits, but this term is not used in the Continuum section or anywhere else in the text; please either use it where the continuum is discussed or remove it.","section":"Glossary"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is explicitly a draft (V0.3, Work in Progress, Request for Comments) and, as a position paper, it may be more suitable for a workshop or a venue with a forum format. The main concern is the gap between the strength of the claims and the evidence: the 'qualitative break' conclusion is stated categorically while the paper's own Unknown Unknowns section says the latency thresholds have not been measured. I would advise the editor to condition acceptance on the authors either substantially tempering the claim to a hypothesis or adding a structured scenario/sensitivity analysis over the plausible parameter range."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a read if you work on agent infrastructure. This is a position paper, not a research result, and once you know that, it is a decent one. The authors know the relevant systems (DNS, PKI, BGP, RDAP, DIDs) and structure the upgrade/switch/hybrid space clearly. The labels—Latent Capability Threshold, AgentFacts—are new wrappers around known ideas, but the comparative table (Table 3) and the discussion of configurable search paths and boundary-aware resolution are genuinely useful for framing the problem. They also do something rare in vision papers: they list their own unknowns. The Unknown Unknowns section openly says latency thresholds 'have yet to be measured' and floats 5/50/500 ms as open questions. That honesty earns credit.\n\nNow the soft spots. The load-bearing premise is the claim that agent requirements are qualitative, not incremental, because agents need millisecond-level lookup and revocation and will number in the trillions. Neither number is established. Worse, the paper itself undercuts the latency premise: Table 1 gives autonomous agents an external goal latency budget of 0.25–3 s (LLM inference plus network). A discovery step inside that budget realistically needs tens to hundreds of milliseconds, not single-digit milliseconds. The Unknown Unknowns section confirms the real threshold could be 500 ms. If that is the case, the argument for a qualitative break weakens substantially; DNS Push, SVCB records, and short-lived OCSP-stapled proofs become viable upgrades, as the authors themselves note.\n\nThe paper also conflates two different quantities: steady-state DNS query latency (10–100 ms) and worst-case update propagation (24–48 h at long-tail ISPs). The 24–48 h figure is real, but presenting it as the representative DNS latency is misleading. The 'trillions of agents' scale is extrapolated from serverless invocation counts and IoT ratios, which is not a workload model.\n\nThat said, the core recommendation—a hybrid of centralized registries for critical agents and federated meshes for the long tail—is reasonable and well argued, as a scenario. The problem is that the paper presents it as a determinate prediction, not a conditional analysis. The authors should either bring workload evidence for the latency and scale assumptions or soften the conclusion to 'if agents require X, then hybrid.'\n\nMy advice: send this to peer review as a position/vision paper. The taxonomy deserves discussion, and the honest framing of open questions makes it useful. But the referee letter should be clear that the qualitative-break conclusion is not currently supported by the evidence in the paper.","headline":"A useful taxonomy and honest draft, but the 'qualitative break' claim is a hypothesis, not a result.","tokens_in":20154,"tokens_out":1396,"would_cite":false,"duration_ms":50846,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that autonomous AI agents impose qualitatively new infrastructure demands—millisecond-level discovery and revocation, cryptographic proofs of behavior—that make today's DNS/PKI/IP assumptions obsolete, and predicts a…","keywords":["Internet of AI Agents","agent registry","DNS limitations","certificate revocation","decentralized identifiers","capability-based addressing","hybrid index architecture","autonomous agents"],"falsifier":"Run a large deployed agent ecosystem and measure the distribution of discovery and revocation times using ordinary DNS with low TTLs and short-lived certificates; if the 99th percentile of both stays under one second without a purpose-built registry, the paper's millisecond-level break is not supported.","tokens_in":1680,"feed_emoji":"🤖","tokens_out":2404,"duration_ms":102392,"temperature":0.7,"pith_summary":"This paper asks whether the human web's infrastructure can be stretched to serve an Internet of AI Agents or whether agents need a purpose-built registry architecture. It argues the break is qualitative, not incremental: agents initiate actions, persist state, migrate between hosts, and spawn helpers, so they need millisecond-level capability lookup and instant credential revocation, while DNS propagation can stretch to 24–48 hours and certificate revocation cannot track trillions of short-lived entities. The paper weighs upgrading DNS, PKI, and IPv6 against clean-slate index designs and concludes that neither pure path wins: a hybrid will emerge, with centralized registries for safety-critical agents and federated meshes for specialized ones. A sympathetic reader should take this as a formative architecture argument that identifies where the legacy stack crosses a threshold and sets measurable targets for whatever replaces it, rather than as a finished protocol proposal.","feed_headline":"Trillions of AI agents will outgrow DNS and PKI","feed_subtitle":"Discovery and revocation must run in milliseconds, so hybrid agent indexes will emerge, the paper predicts.","key_machinery":"The load-bearing object is the agent index or registry: an authoritative system that stores cryptographic identifiers, capability descriptors, trust metadata, and audit logs for agents. The paper positions it as a resolution layer above DNS, IP, and PKI rather than a replacement for transport, and contrasts upgrade paths (SVCB/HTTPS records carrying capability hashes, DNS Push via RFC 8765, RDAP metadata extensions, ACME-plus certificates with short-lived stapled proofs) with switch paths (a clean-slate cryptographic namespace, self-sovereign DID meshes, and capability-first addressing such as '/translate-en-es'). The registry does the work that DNS did for the human web: it makes identity portable when agents migrate hosts, makes revocation immediate, and makes capability claims cryptographically checkable, with the paper's proposed AgentFacts metadata binding an agent's identifier to capability hashes, policy constraints, and runtime attestations.","core_discovery":"On the paper's own terms, the central claim is that the legacy web stack was built for human-initiated, stateless, slowly changing resources and therefore has a latent capability threshold that autonomous agents cross. Crossing that threshold demands three new properties: self-directed discovery (an agent finds and negotiates with unknown peers in milliseconds), delegated authority with instant revocation (helper-agent grants can be yanked the moment they misbehave), and cryptographic proof of behavior rather than proof of domain control. From this, the paper derives that per-agent IP addressing is economically and operationally untenable—$10^{12}$ /32 prefixes would inflate the BGP table by six orders of magnitude—that DNS propagation at 24–48 hours is orders of magnitude too slow for agent churn, and that CRL/OCSP cannot scale to trillions of certificates. The conclusion is that the future is a layered index: centralized registries for high-trust agents and federated meshes such as DID, IPFS, or ActivityPub for the long tail, bridged by gateways that sign and cache proofs so agents can cross tiers in under 50 ms.","pith_inferences":["A consequence the paper leaves implicit is that whoever controls the registry layer inherits the gatekeeper role that domain-name governance and certificate authorities hold today, creating a new contest over which agents get indexed and which capabilities get attested.","A testable extension would be to instrument a production agent framework and measure real discovery and revocation latency envelopes; if 95th-percentile times stay under one second through upgraded DNS and short-lived certificates, the millisecond requirement is an assumption rather than a measured constraint.","If no registry standard is adopted quickly, the likely near-term world is fragmented agent networks separated by bridge gateways, with large platforms running private centralized indexes over federated long tails—mirroring the walled-garden risk the paper flags for naming."],"forward_implications":["Discovery and revocation move from a human-timescale DNS/PKI problem to a millisecond registry problem, with SVCB/HTTPS records and DNS Push as stopgaps rather than endpoints.","Per-agent IP addressing is ruled out, so any workable design routes through an indirection layer that shields BGP from agent churn.","Certificate trust shifts from domain-validated identity to capability attestation and short-lived, stapled proofs, with software-bill-of-materials digests bound to each agent credential.","Enterprises will demand configurable search paths that resolve internal registries first and then public ones, with audit logs for every boundary-crossing lookup.","The likely near-term outcome is hybrid: centralized registries for critical or safety-sensitive agents, federated meshes for specialized ones, and gateway bridges for interoperability."],"supporting_citations":[{"why":"Supplies the current scale baseline of billions of users and smartphones that the paper contrasts with trillions of future agents.","marker":"[1]"},{"why":"Provides the exemplar of a fully autonomous digital agent whose capabilities motivate agent-initiated interaction.","marker":"[2]"},{"why":"Grounds the claim that managed DNS update propagation has consistency anomalies and performance limits.","marker":"[3]"},{"why":"Grounds the 24–48 hour worst-case DNS propagation figure that drives the millisecond-latency argument.","marker":"[4]"},{"why":"Grounds the roughly 44 percent IPv6 adoption figure used to argue that address abundance does not mean usability.","marker":"[5]"},{"why":"Provides the Decentralized Identifiers standard used for the self-sovereign identity switch path and the fragmentation risk of many DID methods.","marker":"[6]"},{"why":"Shows that ACME automates certificate issuance quickly, which the paper uses to isolate revocation as the lingering PKI bottleneck.","marker":"[8]"}],"fun_headline_variants":["Millisecond agents vs 24-hour DNS: hybrid indexes win","AI agents break DNS and PKI; hybrid registry emerges","Trillions of agents need instant revoke and discovery","Why the web stack can't scale to trillions of agents","Upgrade or switch? Hybrid index for AI agent internet"],"cache_read_input_tokens":22016,"weakest_assumption_plain":"The argument rests on unvalidated projections that there will be trillions of always-on agents requiring millisecond-level discovery and revocation; if real agent ecosystems tolerate sub-second or second-level operations, or grow only to millions or billions of entities, the claimed qualitative break with legacy infrastructure weakens.","fun_headline_variants_meta":{"raw":{"variants":["Millisecond agents vs 24-hour DNS: hybrid indexes win","AI agents break DNS and PKI; hybrid registry emerges","Trillions of agents need instant revoke and discovery","Why the web stack can't scale to trillions of agents","Upgrade or switch? Hybrid index for AI agent internet"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1367,"prompt_tokens":974,"completion_tokens":393,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":309}},"tokens_in":590,"tokens_out":393,"duration_ms":6681,"temperature":1.0,"reasoning_tokens":309,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:57:37.864777+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a large deployed agent ecosystem and measure the distribution of discovery and revocation times using ordinary DNS with low TTLs and short-lived certificates; if the 99th percentile of both stays under one second without a purpose-built registry, the paper's millisecond-level break is not supported.","supporting_citations":[{"cited_title":"(2025, March 25)","cited_arxiv_id":null,"evidence_quote":"Grounds the 24–48 hour worst-case DNS propagation figure that drives the millisecond-latency argument."},{"cited_title":"(2025, Apr.)","cited_arxiv_id":null,"evidence_quote":"Provides the Decentralized Identifiers standard used for the self-sovereign identity switch path and the fragmentation risk of many DID methods."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that ACME automates certificate issuance quickly, which the paper uses to isolate revocation as the lingering PKI bottleneck."}],"review_version":1}