# Getting over the nebulosity of agents 06/15/26 No one sent out a memo when the language shifted from "AI" to "agents," but it's everywhere now. That word, specifically, _feels_ a little BS-y — AI itself is undeniable — but the plastering of the word “agent” on every billboard and bus (granted, maybe just in SF) makes one wonder what, specifically, is being described. And how everyone seems to have — already — figured out the sorts of SaaS you can sell for it (to it?). ![[910d568dea4f5718225cde7e371d61fb12a38a2b.mov]] *This* much for a while loop with brains and tools? I'm being coy. I'm running these while loops all day: my coding agents, my Poke and OpenClaw, and more. I like them! ![[IMG_5262.jpeg|My first OpenClaw in January 2026.|300]] I should preface by saying I've [bet](https://www.natural.co/blog/why-krish-joined-natural) the next chapter of my career on “agents.” But even while interviewing: I continuously asked what anyone *meant* by “agent," and how it could be rational to build _so much_ around something so currently indeterminate. I come from [[Ramp#^d91c52|B2B SaaS]], where we relentlessly chase down _real_ customer problems, and solve them with efficiency and craft. Not buzzwords. Is this a real customer yet? I still followed my gut. I know these things start as toys. The team was both sharp and convinced. And sure, we’re building complex, regulated things that require lead time. But why the new word? It’s just AI, right? Has the new term earned its keep? Is there an irrational anthropomorphism behind it? How do you build for a customer who hasn't grown up yet? Is this marketing? Is it delusion? Still, a few months into both my move to SF and the start of the job, I’ve come around (with zeal!) to agents as a distinct phenomenon that form their own category. Anyone who works full-time on agents, harnesses, models (etc) is likely to read the rest of this and go "no shit, sherlock" or "hmm, not exactly." This post is for people like me who — until recently — built conventional kinds of software, yet to start building the new kind full-time. --- AI, the dwarf in the flask — nebula of black dust with a smirk — is real and intelligent. But it has no shape, and it's hard to imagine it as an actor, as opposed to an oracle. It's missing *form.* ![[ChatGPT Image Jun 10, 2026, 07_06_51 PM.png|]] It has forms: eg. ChatGPT, Claude Code, etc. Each *form* is a new force multiplier. But there isn't a *platonic* form that you can unambiguously, reflexively, associate with "agent." There are yet to emerge "homunculi" that inspire speciation. I still _feel_ like Claude Code is — eg. — just a *specific tool* that I can talk to. Kahlil, founder of [Natural](https://natural.co) (where I now work) describes this challenge internally: >*What do you see when you close your eyes and try to visualize an "agent?"* Plato defined man as a featherless, bipedal, animal; Diogenes [rebutted](https://penelope.uchicago.edu/encyclopaedia_romana/greece/hetairai/diogenes.html#:~:text=animal%2C%20bipedal%20and-,featherless,-%2C%20Diogenes%20plucked%20a) with a chicken. ![[Pasted image 20260615154940.png|Sprint planning or Socratic dialogue?|500]] It's helpful to draw analogy (and contrast) to the early aughts of the Internet. A lot of the concepts we now take for granted about *any* web application, were non-obvious at the outset. The web "application" _itself_ — in contrast to a web*site*, primarily informational — was a novelty in its time. You might remember "Web 2.0." The "web" went from read, to read-and-write. Writing used to be a thing you did on your desktop, to files that you owned. The web was for documents. Paul Graham [describes](https://www.paulgraham.com/road.html), in September of 2001: >When we started Viaweb, hardly anyone understood what we meant when we said that the software ran on the server. It was not until Hotmail was launched a year later that people started to get it. Now everyone knows that this is a valid approach. There is a name now for what we were: an Application Service Provider, or ASP. > >I think that a lot of the next generation of software will be written on this model. Even Microsoft, who have the most to lose, seem to see the inevitablity of moving some things off the desktop. It was so unclear at the time that the server-based, browser-based model would dominate computing, that despite *his* success with the model, PG hedges aggressively in the last few sentences of the intro. > **If** software moves off the desktop and onto servers, it will mean a very different world for developers. > > **To the extent** software does move onto servers, what I'm describing here is the future. Even the [idea](https://news.ycombinator.com/item?id=4773859#:~:text=I%20just%20woke%20up%20with%20the%20idea%20that%20we%20might%20be%20able%20to%20control%20the%20software%20on%20the%20server%20by%20clicking%20on%20links.https://news.ycombinator.com/item?id=4773859#:~:text=I%20just%20woke%20up%20with%20the%20idea%20that%20we%20might%20be%20able%20to%20control%20the%20software%20on%20the%20server%20by%20clicking%20on%20links.) to pursue this model (basically) just [came to him in a dream](https://www.fastcompany.com/3002810/aha-moments-made-paul-grahams-y-combinator-possible#:~:text=One%20morning%20Graham,web%20server%20itself). Sometimes the whole game is your willingness to lean into [nebulosity](https://meaningness.com/nebulosity). ![[ChatGPT Image Jun 12, 2026, 12_46_11 PM 1.png|]] The pattern PG foreshadowed here in 2001, would only materialize in the decades that _followed_. The way he was building Viaweb was genuinely weird. He [knew](https://www.paulgraham.com/first.html) it was among the first. The "web" has continually incorporated non-standard eng practices, once considered odd novelties! There is so much "anatomy" we now take for granted. ^861bc1 1. **Users, accounts, auth:** your system having a *particular* relationship with each end-user 2. **APIs (REST, [[Ramp#^b180a1|SOAP]], etc):** presenting _dynamic_ content to end-users 3. **Databases:** _storing_ things that end-users write into your system 4. **Social graphs:** _connecting_ end-users to each other 5. **Payments:** *accepting* money from end-users, paying end-users _out_, letting them _pay each other_ 6. **Interactivity:** from Javascript / jQuery → “single-page apps” → WebAssembly and desktop-grade experiences in the browser, CRDT, WebSockets, SSE, WebGL, etc 7. **Comms:** programmatic email, text 8. **Cloud:** elastic compute, storage, virtualization, etc 9. **(Big) data:** MapReduce, DAGs, streams, logs, NoSQL, key-value DBs, etc 10. **Algorithms & feeds:** delivering highly personalized content to each and every user 11. **Business models:** ads, SaaS, etc. Don't forget: the stitching of all of these primitives to build anything remotely useful. These elements have since faded to the background; they're ingrained in our closed-eyes image of a webapp. Each of these was intensely debated in their time. Motivating many of those debates was a desire to stop letting the server, the browser, JavaScript, the VC-backed startup — any limb of the “web” — do what was once the province of the desktop. There were all kinds of objections — the desktop in some sense was more "yours" than the servers that followed it — but the Internet and web browser ultimately produced better things. I personally remember giving up Sketch for Figma, after a period of belief that Sketch would stay superior. Claude knows how to build a web app, with minimal controversy in the choices it makes across any of these. Does Claude know what you _mean_ when you want it to build an agent? The wager is that _it will_. If you use any sufficiently sophisticated AI tool today — especially the ones that would more explicitly brand themselves as agents — you can already sniff out some of the primitives. 1. **Harnesses:** Claude Code, Codex, OpenCode, OpenClaw, etc 2. **Memory:** "RAG" / embeddings, Git (see eg. [code.storage](https://code.storage) powering Poke), Markdown (eg. Obsidian), Notion, Linear, etc 3. **Model routing:** Open-source models vs. Lab models. System 1, system 2. 4. **Tools:** MCP, CLI, [API](https://x.com/StartupArchive_/status/2066126001948745737/video/1?s=46) (etc) 5. **Skills:** SOPs for agents 6. **Durable execution:** Temporal, DBOS, hand-rolled, etc It hasn't been long enough to feel good about this list, relative to a "platonic" shape — but the shape will consist of things *like* this that prove themselves durably relevant over time. As the list stabilizes, and the mechanisms under the agents we use start to form idioms — whether they're coding agents like Claude Code or Codex, personal assistants like Poke, or others — the concept of an "agent" will ossify. Agents will start to feel both real and anodyne. --- There is a contrast with the Internet that I want to note. The early Internet was highly motivated around open-source, and around composable — networked — primitives that anyone could stitch together freely, to _compose_ their own web-app, network webapps with each other etc. There was a lot of _ecosystem_ work that everyone took really seriously. To the extent anything became more "closed," it was in deviation from the early spirit. It feels like the analogues in "agents" are opaque from the jump; gated by the creators and owners of the most important "agents" (the labs, companies who aspire to _become_ labs, etc). Everyone is constantly freaking out about "moats.” Are the best agents going to be closer to apps and websites? Or will they be tightly-integrated, impossible-to-replicate like an iPhone? This thought came from realizing how *little* I appreciated the details of agents I was already using. Claude Code is not a wrapper over the latest Anthropic model; it has a lot of the _specific_ anatomy described earlier, independent of the raw inference it accesses. I take for granted how my Poke maintains its deep but fluid memory of me & my life (I forced to consider this when I learned that it’s [powered by a nascent Git-for-agents](https://x.com/pierrecomputer/status/2062921701709520897?s=46)). And so on. I’m so used to building the old kind of software, that I forget the still avant-garde possibilities these early examples are mere _demonstrations_ of. But also: there are no longer UI elements, nor user experiences for me to anchor on in terms of architecture. There is no "Inspect Element." Agents, as experienced by people, are deceptively simple faces, attached to novel and unprecedented machinery, that can’t trivially be reverse-engineered by observation. ![[75F8CB52-FDC1-4FBD-A68D-CB26399217EA.png|]] It’s optimal, for the people defending their moats, that you assume everything is a model or a thin model-“wrapper” (derogatory). It’s not remotely true anymore, though. I think this property of agents will nevertheless make it harder for them to become legible conceptually; which slows the extent to which idioms emerge. But I work for an agent infrastructure [company](https://natural.co). We're trying to give _anyone_'s agents — developer, business, consumer, whoever — both wallets & payments. That implies a strong belief — mine, the company’s, and capital’s — in the proliferation of agents. There are a hundred billboards in my SF neighborhood trying to sell me, a random developer, infra for _my_ agents. So maybe market incentives will still take care of things. As I write this, Satya Nadella (Microsoft CEO) published [*"A frontier without an ecosystem is not stable"*](https://x.com/satyanadella/status/2066182223213293753): >Companies need to turn their workflows, domain knowledge, and accumulated judgment into AI systems that improve with each use. > This loop becomes the new IP of the firm. > >In my view, our priority has to be building a frontier ecosystem, not just a frontier model, so value flows broadly across every company, every industry, and every country. One where every organization can own the learning loop that encodes its institutional knowledge, compounding its human and token capital. Satya is pitching an ecosystem where *every business* can [[Ramp#^9f3029|build their own]] agent, which: 1. implies a future with _many_ agents 2. for most businesses, would require that the *primitives* (harnesses, memory, tools) exist in an open ecosystem accessible to developers Microsoft has every incentive to say this. But I do think that when it comes to the application of "intelligence," we will find that [God hates singletons](https://www.palladiummag.com/2023/12/01/god-hates-singletons/). A *bunch* of agents running around is actually a relief, compared to the “SkyNet” we’re primed to expect. ![[ChatGPT Image Jun 15, 2026, 07_02_24 PM.png|]] My work and life already relies on just a handful of “agents”. I can't stop noticing how _incomplete_ they are, how little competition there is, and how many of their moving parts need to get better. Even just [[Getting over the nebulosity of agents#^861bc1|recapping the Internet]] makes me realize how long a road like this stretches. There are _decades_ of work left in order for agents to be good and effective. But they’re real. Billboard away!