Who we follow.

Azhar runs his entire chief-of-staff operation — "R Mini Arnold," an OpenClaw agent fleet — on a single Mac mini in his garden studio: the local-first, own-your-hardware sovereignty bet Tropo is built on, lived at the frontier. And Exponential View, read by hundreds of thousands, named identity, verifiability, and coordination as the invariants of the agent economy — three of the axes Tropo composes into one governed whole.

Swyx coined "AI Engineer" as a discipline and, in "Scaling without Slop" (Jan 2026), argued the defensible move in an age of effortless mediocrity is not more output but better quality-at-scale through editorial judgment, curation, and evals — "changing the slope of slop, not giving up on humans." That is verification-as-moat in the builder mainstream's own words, with a handle the whole field now uses.

Orosz's 2026 "slow down to speed up when working with AI agents" names the exact tension: agents now ship 2x+ the code, so the constraint moves from writing to verifying — making tests, observability, and evals the real engineering moat rather than a tax. The Pragmatic Engineer is where that case reaches more than a million working engineers, sponsorship-free.

In "After Automation" (May 2026) Shipper argues that as AI commoditizes routine competence, every agent needs a human to point it at the right thing and judge whether the output is good — the AI paradox is that automation creates more expert human work, not less. That bounded-verification stance is verification-as-moat, said by the CEO of a company that actually runs on agents.

In "The Inference Shift" (2026) Thompson drew the distinction between answer inference and agentic inference, arguing the latter is "less about GPUs answering a question and more about the memory hierarchy wrapped around a model." That is the bet Tropo makes from the builder's seat — the durable, governed context layer is where value compounds — named at the strategy altitude through his Aggregation Theory lens.

In "The Shape of the Thing" (Mar 2026) and "Management as AI superpower" (Jan 2026) Mollick framed the era as one of "managing AIs, rather than working with them" — you hand a fleet of agents hours of work and supervise the outcomes. That fleet-as-a-team thesis, from Wharton's Co-Director of the Generative AI Labs, is exactly the governed-agent surface Tropo turns into infrastructure.

Gupta's "AI Product Operating Model" (with Rohan Varma) names the premise Tropo is built on — when building code becomes cheap, AI-native orgs invert to build-first and collapse coordination layers. He gestures at taste and verification as the new scarce inputs, and his "PM's Guide to Agent Distribution: MCP, CLIs, and AGENTS.md" lands squarely on markdown-as-substrate.

Jones's most-loved piece, "Why your AI starts from zero every time you open a new chat," is the durable-memory thesis for practitioners, and "cheap intelligence is not the same as being able to use it" is verification-as-moat in one line. His June 2026 briefing "Your team is running agents nobody owns" is the agent-ownership problem a governed studio answers — he teaches one rung below where Tropo governs.

In "Why I don't think AGI is right around the corner" Patel named the real gap: agents have no continual learning — no way to consolidate experience over time, the missing "sleep" of memory reconsolidation. He framed the problem the whole field now debates; Tropo's governed memory and agent succession are the present-day, local-first answer to the gap he diagnosed.

In "The natural home for AI agents is your Reminders app" (Jan 2026) Webb described agents that "sequence tasks into steps, and pause where clarification is needed or, for trust reasons, you want a human in the loop" — citing Claude Code's permission scoping and asking for "progress bars not notifications." That governed, legible coordination layer for a team of semi-autonomous agents is the core of what Tropo builds.

Wardley's mapping method shows the model and harness tiers industrializing into utilities while durable value migrates up to the institutional layer that governs and orchestrates the commoditized parts — "don't bet on the latest technology; understand when to exploit industrialization and when to invest in genesis." That is the strategic case for Tropo's bet: a governed, owner-controlled OS above commoditizing agents.

Lambert coined RLVR — Reinforcement Learning with Verifiable Rewards — in the Tülu 3 paper, making checkable correctness a first-class training signal, and paired it with the line that "being good at using AI today is a better moat than working hard." Interconnects is where the verification thesis gets stated from the frontier-research side, the same axis Tropo builds its substrate on.

Karpathy's April 2026 LLM Wiki is almost line-for-line how Tropo treats memory: raw sources distilled into an LLM-maintained, schema-governed, version-controlled markdown wiki. He didn't just bless markdown-as-substrate — he surfaced the exact open problems a governed vault exists to solve (multi-writer consistency, hallucination accrual, scale). Now leading pre-training research at Anthropic, the same Claude ecosystem Tropo builds on.

Howard's llms.txt proposal (2024) made the strongest public case that the right substrate for machine-readable knowledge is markdown a human can also read — a curated index over clean .md, built so classical parsers and regex still work. He helped catalyze the broader markdown-as-protocol wave; Tropo is what that single-file idea looks like grown into a typed, validated, governed institution.

Litt's "Code like a surgeon" frame — the human stays the high-leverage operator doing the real work while a prepped team of agents handles setup — is the most precise articulation of how Tropo treats agents: a managed crew with clear lanes, not autonomous replacements. His decade of malleable-software and local-first work (software you own, shape, and run yourself) is the axis Tropo extends into the agent era.

Willison's "lethal trifecta" — private data, exposure to untrusted content, and the ability to communicate externally — is the cleanest case anyone has made that trustworthy agents require governance and verification by construction, not bolted on after. His markdown-as-source, own-your-tooling workflows and the LLM CLI are Tropo's governed vault in miniature, built from the practitioner's bench.

Husain dragged "evals" out of research jargon and into working practice — systematic error analysis and LLM-as-a-judge over vibe checks — and co-published the practitioner guide to LLM evals in The Pragmatic Engineer with Gergely Orosz. His whole thesis is that measurement is what separates a demo from a product you can trust: verification-as-moat, taught hands-on to thousands of engineers.

Horthy's "12-Factor Agents" reframed reliable agents as mostly-deterministic software with LLM steps placed on purpose — own your prompts, own your context window, contact humans via tool calls. His RPI / spec-first workflow treats the spec as the durable contract agents code against. That is verification-and-discipline as the moat, the same instinct Tropo builds into governed, plain-file substrate.

In "Don't Build Multi-Agents" Yan showed that parallel agents quietly diverge because every action carries an implicit decision the others never see — his single-writer principle insists shared context be one durable thing, written down, not re-implied in each agent's head. An early, influential proponent of context engineering, arguing from inside a frontier coding lab the move Tropo makes: decisions belong in the substrate.

In "Your harness, your memory" (Apr 2026) and Deep Agents, Chase made the agent's harness and its filesystem the load-bearing primitives — the agent works against files in a workspace, not just a context window — and on the Sequoia podcast argued context engineering, not the model, is the moat. The incumbent framework is converging on files-as-substrate from the opposite direction; Tropo governs them.

Moura built CrewAI's role-based crew model and, in "Stop giving your agents database credentials" (Jun 2026), argued the platform — not the agent — must own permissions, data boundaries, secrets, and telemetry: "the agent should go through the same governance layer your analysts go through." That governed-boundary instinct is exactly what Tropo formalizes for a standing team of agents.

Fanelli's "Can coding agents self-improve?" found that GPT-5 and Claude Opus would build useful dev tools and then refuse to use them ("I didn't need any of them") — landing on the point that prompting alone won't make agents adopt their own improvements; you need a harness that holds them to it. As co-host of Latent Space and builder of AWT, he lives the verification-and-tooling thesis Tropo runs on.

At Cursor and now Codex, Varma watched coordination overhead collapse once code got cheap and put his finger on where the scarce work went — "the taste to know which built version is actually right." His "build first, evaluate second" thesis is verification-as-moat said by an operator, not a theorist; Tropo exists to give that judgment a substrate.

Steinberger compressed the shift into one line — "you shouldn't be prompting coding agents anymore; you should be designing loops that prompt your agents" — a June 2026 post Latent Space canonized as "Loopcraft." His OpenClaw (the most-starred project in GitHub history) is the loudest proof that local-first, own-your-data agent demand is mainstream — the same sovereignty bet under Tropo's studio.