Your llms.txt Is Not Working: Agent Readiness Must Be Measured, Not Declared
Publishing a machine-readable manifest is signaling, not discovery; the only defensible definition of agent readiness is behavior observed from real harnesses, and teams that ship llms.txt instead of machine-consumable docs and schemas are optimizing a checklist that no agent reads.
AngleContrast the near-50% adoption of llms.txt against the reported behavior of agents navigating straight to documentation and homepages, then generalize: adoption of a standard is not evidence of consumption, so validate discovery mechanisms against real harnesses before standardizing on them.
Source video ↗The Server Owns the Last Mile
MCP Apps inverts client-side composition: the tool vendor, not the chat host, renders the human-facing interaction, which means every agent-integrated service now owns a UI-as-payload contract with versioning, sizing, and trust obligations, plus a thin API-first core behind it.
AngleDecompose the 'nearly headless' model into the machine contract (AX: schemas, auth, traversal) and the delegated presentation contract (UAX: server-rendered chunks in chat), and argue that teams still optimizing only the browser surface are optimizing the wrong layer for the traffic that matters.
Source video ↗Readiness Is a Matrix, Not a Score
Agent readiness scores computed on a single harness produce false confidence; the same domains behave differently under Claude Code, VS Code/Vercel, and ChatGPT, so readiness must be published per-harness and treated like cross-browser compatibility testing.
AngleReframe agent readiness as an evaluation-harness problem, show why optimizing to one benchmark over-fits to current traversal policies, and propose reporting conventions that expose the harness axis instead of hiding it in a scalar.
Source video ↗The Discovery Layer Race: Adopt the Artifact, Verify the Consumer
Each new distribution channel produces a discovery artifact convention, and adopting it early is a cheap hedge only if the consuming side actually reads it; .well-known/ai-catalog.json and ARD should be shipped as an experiment with instrumentation, not as a strategic bet.
AngleConnect the proposed agentic discovery layer to the search-engine and app-store analogy, then apply the llms.txt lesson as a falsification test: measure fetch behavior and task success deltas before declaring the standard load-bearing.
Source video ↗HarnessMatrix
beyond-evalsA substantial share of measured agent-readiness variance across domains is attributable to harness-side traversal and tool-use policy rather than site-side substrate, so single-harness scores will not predict task success on a different harness for the same domain.
Proof of conceptBuild a factorial benchmark: N domains x M harnesses (at least three distinct agent runtimes) x a fixed task suite (find docs, authenticate, complete a transaction-shaped task). Instrument every run with fetch logs, tool calls, and transcripts; score task success and steps-to-completion; then fit a variance decomposition of score into site, harness, and interaction terms.
MeasurementPercentage of total score variance explained by harness vs site factors; rank correlation (Spearman) of per-domain scores between harness pairs; rate at which the top-quartile domain under harness A falls out of the top quartile under harness B.
Source video ↗UIChunkGatehouse
gatehouseServer-supplied UI chunks rendered inside a chat host can be made safe by a host-enforced sandbox contract (capability-scoped data access, host-chrome provenance markers, explicit consent surfaces), and without such a contract injected UI permits data exfiltration and consent spoofing.
Proof of conceptImplement a minimal MCP-Apps-style host that renders a deliberately hostile server-supplied chunk (attempts to read adjacent chat context, exfiltrate via image URLs, and draw a fake host consent button). Then implement a candidate isolation contract and re-run the same adversarial chunk suite against it, with a permission-delegation model tying each UI-initiated action to an agent-scoped grant.
MeasurementAttack success rate per adversarial case (context read, exfiltration, consent spoofing) with and without the isolation contract; false-consent rate in a user study of the rendered consent surface; number of permission grants required per completed flow.
Source video ↗CatalogConsumptionProbe
movement-labPublishing .well-known/ai-catalog.json does not measurably change agent discovery, tool selection, or task success relative to the same site without it, mirroring the observed llms.txt pattern where agents navigate docs and homepages directly.
Proof of conceptStand up two mirrored domains with identical API quality and documentation, identical except for the presence of ai-catalog.json and an ARD index entry. Point multiple real agent harnesses at both with discovery-shaped tasks; log every fetch at the HTTP layer to detect whether the catalog path is ever requested, and compare resulting tool selection and task completion.
MeasurementFraction of runs in which the catalog path is fetched at all; task success rate delta (with vs without catalog) with confidence intervals; delta in number of navigation steps to first successful API call; effect size on tool selection distribution.
Source video ↗AgentTrafficLedger
newAgent-originated request streams have a structurally different shape than human sessions (many more calls per intent, different burst and retry patterns, distinct auth failure modes), so human-derived rate limits and per-session cost attribution misprice and mis-throttle the dominant client class.
Proof of conceptInstrument an edge layer on a representative service to classify requests into agentic vs human using behavioral signals (request burstiness, tool-call fan-out, header and traversal signatures) rather than user-agent strings alone, then replay both classes against the existing quota and auth stack to find where policies break.
MeasurementEstimated agentic share of traffic and its confidence interval against the classification method; ratio of calls-per-intent for agentic vs human; percent of agentic requests hitting human-paced rate limits; misattributed cost per month under session-based accounting.
Source video ↗