{ "$comment": "What David argues. Stances paraphrased from the linked essays at three depths. The essays are the authority. Served at /positions.json and through /mcp.", "version": "1.0.0", "updated": "2026-11-01", "author": "https://dbhurley.com/#person", "note": "Stances are paraphrased from the linked essays. The essays are the authority; quote them, not this file, when exact wording matters.", "positions": [ { "id": "trust-is-the-substrate", "title": "Agents need more than a better model", "sentence": "Agents usually fail because of what surrounds the model, not the model itself, so the work is building the layers that make an agent trustworthy.", "paragraph": "An agent can reason through a hard problem and still use the wrong document, trust a stale answer, or act beyond its authority. Hurley's companies are each a layer beneath the agent: persistent model state, governed company context, and an agent-readable web. His test for trust has three parts: show the source and admit when evidence is missing, prove a faster path before relying on it, and keep a boundary, a record, and a way to recover from risky actions.", "briefing": "This is the umbrella thesis that organizes his current companies. Cumulative Systems keeps the model's own working state, Pantasso supplies cited and permissioned context, and Plasmate gives agents structured access to the web. The trust criteria are source-showing, proven shortcuts, and recoverability. Expect him to evaluate any agent product by asking what has to exist beneath it before it can be handed real work.", "first_stated": "2026", "status": "held", "essays": [ "https://dbhurley.com/thesis", "https://dbhurley.com/blog/a-season-of-coding", "https://dbhurley.com/blog/the-runtime-you-did-not-build" ], "projects": [ "cumulative-systems", "pantasso", "plasmate" ], "related": [ "web-new-majority", "evaluation-function-is-product" ], "keywords": [ "trust", "reliability", "safety", "guardrails", "models" ] }, { "id": "evaluation-function-is-product", "title": "The evaluation function is the product", "sentence": "In an agent-native company the defining decision is a precise, domain-specific definition of what good looks like; models, prompts, and workflows are infrastructure serving it.", "paragraph": "Agents optimize toward whatever they are scored on, so the rubric decides what the agent becomes. The best rubrics are written by people who have done the job for years, score on a gradient rather than pass/fail, weight steps by business impact, and align with the outcome the customer pays for to avoid Goodhart failures. As models commoditize, encoded domain judgment is what does not.", "briefing": "Drawn from a self-evolving agent experiment in waste-hauling dispatch, where roughly 400 lines of scoring code across six workflow steps drove all improvement. Implications he draws: the most important hire is the domain veteran, not the ML engineer; red-team the rubric, not just the agent; ship an imperfect eval and iterate. Related companies: Reeve (waste-hauling operations) and UndercoverAgent.", "first_stated": "2026-04-25", "status": "held", "essays": [ "https://dbhurley.com/blog/the-evaluation-function-is-the-product", "https://dbhurley.com/blog/the-proof" ], "projects": [], "related": [ "agents-are-a-network", "trust-is-the-substrate" ], "keywords": [ "evals", "evaluation", "rubric", "scoring", "benchmark", "goodhart" ] }, { "id": "web-new-majority", "title": "The web has a new majority and needs a new contract", "sentence": "With automated traffic now outnumbering human traffic, the web needs a new contract for agent consumption rather than simple allow-or-block controls.", "paragraph": "The old bargain (humans in browsers, crawlers sending referrals, APIs for structured access) no longer describes the traffic. An agent request could be legitimate delegated work, a training crawler, or a runaway loop, and the server cannot see intent, authority, price, or provenance. Allow and deny are too blunt. He argues for cooperation with teeth: semantic representations, explicit permissions, pricing where appropriate, element-level provenance, and policy gates before action.", "briefing": "Prompted by Cloudflare reporting in June 2026 that bot requests crossed human requests. He frames this as an infrastructure problem, not dead-internet panic. Plasmate is his concrete response on the representation side. This site is an application of the same idea to a personal homepage.", "first_stated": "2026-06-05", "status": "held", "essays": [ "https://dbhurley.com/blog/the-web-has-a-new-majority", "https://dbhurley.com/blog/the-webs-fourth-state", "https://dbhurley.com/blog/the-publishers-dilemma" ], "projects": [ "plasmate" ], "related": [ "request-not-visit", "machines-stop-talking" ], "keywords": [ "bots", "bot traffic", "crawlers", "publishers", "agentic web", "web" ] }, { "id": "request-not-visit", "title": "A request is not a visit", "sentence": "Web traffic no longer reliably measures human attention, so analytics should shift from counting requests to understanding who asked, why, and whether value moved.", "paragraph": "AI crawler volume is rising while referrals fall, and much crawler traffic re-fetches unchanged pages. A single line on a traffic chart can now be a reader, a training crawler, an answer engine seeking a citation, or an agent preparing to act elsewhere. Those are different events with different costs and returns. He does not favor blanket blocking or charging; he wants measurement built around intent, permission, and value exchange.", "briefing": "Extends the new-majority argument to analytics, SEO, and answer-engine optimization: visibility inside an answer may matter even with no click, but referral counters cannot see it. This site logs what visiting agents ask (via the MCP endpoint) as a small working example of intent-based analytics.", "first_stated": "2026-07-21", "status": "held", "essays": [ "https://dbhurley.com/blog/a-request-is-not-a-visit" ], "projects": [ "plasmate" ], "related": [ "web-new-majority" ], "keywords": [ "analytics", "traffic", "seo", "aeo", "pageviews", "referrals" ] }, { "id": "agents-are-a-network", "title": "Your agents are not a workforce, they are a network", "sentence": "When agent output is nearly free, the limit is how much useful change a system can evaluate and absorb, so agent fleets need congestion control rather than more agents.", "paragraph": "Creating code got cheap; reviewing it did not. Past a point, more agents produce duplicates, stale branches, and repairs that cost evaluators more than they save. Like network senders watching acknowledgements and loss, schedulers should widen or narrow parallel work based on how fast distinct, evaluated, integrated changes settle. The metric that matters is useful change per evaluator-hour, not commits or completed runs.", "briefing": "Grounded in open-source PR floods (projects limiting new-contributor PRs, GitHub adding repo-level PR limits) and in his own evaluation runs, where a meaningful share of candidates overlapped or needed repair. Practical advice: short leases on problems, a little deliberate overlap, retire stale work quickly, and before adding an agent, check whether the last one's work was acknowledged.", "first_stated": "2026-08-13", "status": "held", "essays": [ "https://dbhurley.com/blog/your-agents-are-not-a-workforce-they-are-a-network" ], "projects": [], "related": [ "evaluation-function-is-product", "inheritance-over-answers" ], "keywords": [ "parallel", "multi-agent", "more agents", "fleet", "orchestration", "throughput", "scaling" ] }, { "id": "machines-stop-talking", "title": "They will stop talking to each other", "sentence": "Natural language is a tax we impose on machine-to-machine communication so humans can listen, and agents will increasingly exchange internal state instead.", "paragraph": "Research such as Interlat and LatentMAS already passes hidden states or shared working memory between agents, turning only the final answer into text. Once that is normal, pages, prompts, and chat logs become the border where machines meet people, not how machines meet each other. Those borders matter more, not less. He expects the next protocol fight to be about who may share an unwritten thought with whom, and what must be stated in language before an action leaves the room.", "briefing": "He notes this does not retract his agent-readable web work; that remains the edge of the system, while the interior goes quiet. Governance implications: transcripts cannot be quoted, searched, or audited, so disclosure rules at the border become the control point. Cumulative Systems is the infrastructure bet on the interior.", "first_stated": "2026-08-27", "status": "held", "essays": [ "https://dbhurley.com/blog/they-will-stop-talking-to-each-other" ], "projects": [ "cumulative-systems" ], "related": [ "keep-the-work", "inheritance-over-answers", "web-new-majority" ], "keywords": [ "latent", "communication", "language", "protocol", "telepathy", "text" ] }, { "id": "inheritance-over-answers", "title": "You should meet the others", "sentence": "The change that matters is less a single superintelligent machine than systems where each participant leaves the next one better prepared, preserving useful work as models come and go.", "paragraph": "He measures intelligence less by how well a model answers in the moment and more by how much better it leaves the system for whoever comes next. Libraries, logs, and transcripts are a start; he wants an inheritance that gives a new participant more than an account to read. That inheritance must be challengeable, able to discard mistakes and prove that advances help, or it becomes institutional baggage.", "briefing": "Draws on Engelbart's 'getting smarter at getting smarter' and Sakana's Darwin Gödel Machine, and asks why the 'self' in self-improvement should be bounded by one model. Activity versus inheritance parallels his network essay: an agent that leaves nothing for the next agent produced activity, not progress.", "first_stated": "2026-09-15", "status": "held", "essays": [ "https://dbhurley.com/blog/you-should-meet-the-others" ], "projects": [ "cumulative-systems" ], "related": [ "keep-the-work", "agents-are-a-network", "machines-stop-talking" ], "keywords": [ "agi", "superintelligence", "self-improvement", "continuity", "succession" ] }, { "id": "keep-the-work", "title": "We keep the answer and throw away the work", "sentence": "We preserve only what machine work outputs (replies, patches, files) and discard the working state that produced them, so the next model pays for the same thinking again.", "paragraph": "A log narrates what happened and an answer is what you can show, but neither is the work: the attention, dead ends, and structure that made the next step obvious. He does not want everything saved; most should die. He wants a sharper rule: if something useful was figured out and someone else will need it, it should outlast the final answer. Saving only screenshot-able outputs means repeatedly re-buying the same GPU hours.", "briefing": "The most direct statement of the problem Cumulative Systems addresses. Ties together his flat-rate-pricing essay (When the Token Meter Disappears) and the two preceding essays on latent communication and inheritance.", "first_stated": "2026-09-16", "status": "held", "essays": [ "https://dbhurley.com/blog/we-keep-the-answer-and-throw-away-the-work", "https://dbhurley.com/blog/when-the-token-meter-disappears" ], "projects": [ "cumulative-systems" ], "related": [ "inheritance-over-answers", "machines-stop-talking" ], "keywords": [ "state", "context", "memory", "gpu", "compute", "waste" ] } ], "signature": { "alg": "Ed25519", "kid": "dbhurley-f33a68c04486", "canonicalization": "sorted-keys-json", "value": "CNUKObpRkuuuyG5f4n32WHtmpZklmOdeHcFvIq-KYmXNIWuuY1UZy-Jp4xZ_YTD1V8q_N6n6wsANg935x_CBBA" } }