The More You Constrain an Agent, the More You Can Trust It

Nathan Labenz opens by noting how much the word *agent* has slid around since GPT-4. Projects such as AutoGPT and BabyAGI sold a picture in which you handed a model a grand goal and it planned, browsed, and executed on its own. A year and a half later that path has not quite worked, even if it still might. In the meantime every vendor has fudged the term, which is why he wants Andrei Oprisan, engineering lead at Agent.ai, to say what *he* means when he talks about building agents.

Andrei treats the question as foundational and temporary. It is early, he says, and the definition will keep changing every three to six months; six months from now they will tweak it again. For now he uses a restrained formula: a semi-autonomous system that can perform tasks or make decisions on behalf of a user. The important half of that sentence is the constraint. He wants agents to do very small, concrete, extremely bounded jobs, not the open-ended goals of the first wave.

The future he describes is not one super-agent but dozens or even hundreds of specialized ones sitting next to a knowledge worker. One reads morning email and re-ranks it by preferences that change from day to day. Another looks at the calendar and explains who you are meeting — Nathan for this podcast, Dharmesh for an Agent.ai strategy session — with context that drifts over weeks. Abstraction that decides which agent gets which slice of the job can come later. Overselling autonomy, he warns, is a recipe for hallucination if not for disappointment altogether.

Nathan draws a spectrum that AI, in his phrase, destroys as a clean binary. At zero agency sits a fully deterministic workflow with no model at all: pure code. At the other end sits something like AGI, told to go make money on the internet and trusted to come back with cash. A lot of failure, he thinks, is people not knowing where they should sit. He cites a previous guest who built an AI sales-development rep as a pipeline of roughly 16 steps, each deterministic, with the model never trusted to choose a fork. That design, he says, already works.

Andrei does not think the industry can leave that highly prescribed style yet. Algorithmically and in terms of language-model evolution, he says, we are a few generations away from dropping it, and that is where he expects everyone to live for the next six to 24 months. He flags the then-upcoming ChatGPT model as a possible change on planning, but he is not designing as if the gap is closed. Language models, he says, are fairly bad at planning and worse at helping a person see where the uncertainty is. They answer with confidence. They do not know how confident they ought to be, or when they should doubt themselves.

Two other missing pieces sit beside planning. One is robust out-of-domain detection: noticing when a marketing task has quietly become a sales task, or when the work has left the domain the prompts were written for. The other is reliable self-correction, which first requires a picture of ground truth — what a right answer even looks like — and then a way to steer back toward it. Without that, the model will veer left for fifty steps and narrate the detour as success. Benchmarking is the unglamorous fourth piece. What "really good" looks like for Nathan and for Andrei, using the same agent, can be entirely subjective, so the system needs a way for a user to say they want more of *this*, to explain which of ten or fifty decisions produced it, to reproduce it, and to run on the order of a million simulations until the path tacks closer to something you can describe mathematically.

Those three capabilities — planning, out-of-domain detection, and self-correction — are, in his view, what you need before anything resembling autonomy. The research is not new. Yann LeCun, Meta's chief AI scientist and formerly an NYU professor, has been talking about error correction and domain boundaries for years; Andrei remembers being a student in that class some twenty years ago. His own machine-learning background was building reliable self-correction algorithms: defining ground truth, then doing "surgery" on a neural network at runtime instead of retraining a GPT-4-class model. There is progress in highly vertical domains. Generic planning that does not collapse into a workflow builder with a predefined menu is still very hard. He would love to be pleasantly surprised by the next model. He is building as if those holes remain.

The lesson from Agent.ai's own catalog is blunt. As they built more agents, they learned that the more prescriptive they were, the better the agents performed — in NPS surveys and in concrete user feedback. The flagship example is a company-research agent: enter any company name or domain and it returns a full report, from founding story and decision-makers' emails to LinkedIn and X traffic, product offerings, price points, and a sketch of how you would sell into that account. That one agent has about 20 sections and more than 100 steps, most of them hardcoded. Structured outputs are what let them give the model a little creative room without losing the ability to quantify, iterate, and mix models. Users can then ask, more consistently, who the top three competitors are and how you would compete.

The product bet underneath is a WYSIWYG workflow toolkit plus language-model actions that snap to structured outputs. An expert can drag and drop. Everyone else can describe the job in thirty seconds — read this podcast, summarize the learnings, lean on the topics I care about — and, with some tweaking, get an agent that does that concrete task well. Broader, less structured asks fail more often. Those failures are the MVPs. Mature versions, Andrei says, will cross domains only if the system can tell when it has left the domain: podcasts are not YouTube videos are not email, and the prompts, the few-shot data, and the fine-tunes that buy another five or ten points of accuracy all differ.

Nathan notices one place where Agent.ai already trusts the model more than classic no-code tools do. Filter and routing steps on those platforms are usually exact match, contains, or a regular expression. On Agent.ai the decision can be a language-model judgment over unstructured or semi-structured data: advance under this gist, stop under that one. Andrei's guidance is that for complex if-then-else — and soon looping, retries, fifty ad-copy variants, ten emails to A/B test — the LLM is actually very good. Human beings are not great at enumerating every path a process might take. Even a two-level nest, a while loop, a for loop, and a handful of sentinel words is hard for an expert who lives in the builder every day to wire so it matches the picture in his head. He can explain the intended behavior in conversation, and the model does a better job than he does. They have benchmarked it. It is true for this class of logic. Deeper planning is still out of reach.

Most of the agents on the site today are still hand-built: take this user input, fetch that source, prompt a model, force a structured object, render the result. The same Smart Action Builder can also be driven in chat. Ask it to collect a YouTube URL and it fills in the heading, the variable name, how many videos to crawl, whether the target is a video or a channel, which topics matter. Internal benchmarks put that generation at about 80 percent good — not yet good enough to open widely. Templates are the next lever. Starting in September, Agent.ai planned to publish example agents in marketing, sales, and customer service, either closed for use-as-is or open so anyone can copy them, add three steps, remove two, and email the output instead of showing it on a page. With thousands of those templates as ground truth, Andrei thinks the builder can reach 90 percent-plus accuracy.

The longer vision is a marketplace with App Store economics. The platform would guarantee distribution — he mentions 100,000 eyeballs a month on an agent as a free default — plus a review process so agents do what they claim and do not siphon personal data. Pricing is still an experiment: a flat fee, a rev-share on credits, a 99-cent habit, a $5 utility, or a ChatGPT-like $20 bundle with your own data and a library of agents. Some agents will have a free tier and a pro tier that connects more expensive sources. They have a small alpha of builders, not quite beta, and they intend to seed the catalog with dozens of agents, then let market dynamics take over. As of earlier that week they were past 40,000 users and growing fast; a few minutes later he puts the project at three to five months old and close to 50,000. You should not need to be a developer. Citizen AI builders, in his phrase, should describe what they need the way they already build a form, or pick from examples, and go from zero to expert in weeks if they are willing to fail fast.

Nathan's own mantra is copy and customize. When he saw the Clone button on published agents, he recognized the same philosophy. The header of the site already puts Agent Network beside Human Network. Andrei wants you to add agents to a team the way you add people: a social-media crew, a marketing crew, a set of specialists looking up MBA case studies. If work stays hybrid, agents will be part of it, and a professional social network is how you inspect what an agent can do, what employers say about it, whether it has endorsements or something like certification. Point solutions will keep existing. He does not think they will win long-term adoption once the underlying models start blending domains and someone has to coordinate across the top.

The workflow-builder is the most performance-critical model they run, so Nathan presses on what actually moved the number: explicit instructions, few-shot examples, RAG, or fine-tuning. RAG, Andrei says, has worked surprisingly well. They first wrote a few hundred examples from very simple to very complex, a template library, then a benchmarking tool of maybe 100 or 200 lines of code: here is the goal, here is the agent we would choose, here is why this ask needs YouTube data and Google News. Fine-tuning the same job produced more false positives. Given a fairly predictable set of intents, RAG on larger models did better. He is careful not to universalize. Change the data and the answer changes. What is not optional is an evaluation loop that can tell you, even when the answer looks right, whether you would have known if it were wrong. Another model call, a few good and bad examples, and a JSON enum of good, bad, or I don't know is enough to start — if you keep babysitting temperature, context, and examples, and if you do not pin yourself to a floating "GPT-4 latest" checkpoint that silently moves.

RAG still has knobs that bite. Document size and overlap have to be tuned; a window that is too small, with too little overlap, caused a lot of issues. Their documents are well defined because the objects are: an average agent has six steps, each step a JSON object with six attributes, so they can size chunks to be just big enough. The embedding target is not the proprietary structure. It is a natural-language description, down to the field, including prerequisites — this step needs a YouTube URL; if you do not have one, you can search the web or look up a channel. Retrieval hits the paragraph, which then points at the structured config. Nathan translates that as: do not embed the API blob unless you are ready to train your own embedding model. Andrei agrees, and treats the mapping as data pre-processing. Domain-specific vocab can justify a custom embedding model, but a 2 percent lift is often not worth the complexity. For most of their cases, a tailored prompt with as much context as fits is good enough until the footprint will not fit.

Compression was a pleasant surprise. Ask the model to compress a document, attach a definition with every chunk, and a Claude 200k-token window has carried the equivalent of about three million tokens at the same speed. There are papers on this now, and smaller models can do the same trick. They started those builder experiments on small on-device models for more consistent test answers, then worked back up to Claude's larger context for personalization, retuning prompts and RAG because the same recipe does not transfer. Something that can run on a device, Andrei says, is likely good enough for 99 percent of these evaluation use cases. The project is a few months old. Users should not have to try twelve models and live in the papers to get a useful agent.

Nathan's usual advice is the opposite of a local-first loop: ignore cost at the start, use the absolute top model, get the task to work, then hunt for efficiency. Andrei agrees with the starting point. If you cannot get good answers out of GPT-4, Claude, or Gemini, you probably have the wrong prompt or the wrong data in the window. Hosting your own model is worth it in regulated settings, and even then hospitals in Boston — one of his MBA classmates among them — use ChatGPT under enterprise agreements that promise the data will not be used for training. Cost will keep falling; he expects another order-of-magnitude drop. Think hard, he says, before you decide you need something custom.

Local models earn their keep in a different way. He runs a local embedding model over his personal documents and images so financial records and tax files never go over the wire. Small open weights can be patched and fine-tuned on a machine you hold; you cannot do that to GPT-4's parameters. For a document-definition task he needed to tweak, a local fine-tune beat GPT-3.5; then GPT-4 fine-tuning in the cloud, once it shipped, beat the local model "out of the water." Claude still had no fine-tuning. On daily work he gets results from Phi running locally that are as good as GPT-4, and on some planning benchmarks Phi scores in the fifties — still abysmally low — against the twenties or thirties for Gemini. Smaller nets, he argues, have less pretrained data sitting in the way of what you actually want the model to do. For 99 percent of people that stack is not worth the complexity. Apple Intelligence on-device, plus a private cloud that even Apple cannot read, is the hybrid he expects to spread.

On fine-tuning itself, Nathan has found huge value in chain-of-thought traces that show how he wants the system to reason. Andrei distinguishes that from chain-of-retrieval across multi-step tool use, which he has found harder and more infrastructure-heavy. For chain-of-thought, the data is not just outputs. It is intern-level intermediate steps, highly curated, then supervised fine-tuning on those annotations. He writes prompts that ask the model to break the problem into steps, then evolutions of that prompt to refine the examples; in math and legal reasoning the accuracy gains are measurable. A meta-trick he likes is asking how the task can be broken down further and which decision points they have not discussed. More than once that question revealed they were implicitly joining two concepts that should be split, and splitting them raised accuracy. Nathan's seed-set advice from Waymark, which generates marketing videos for small businesses, rhymes with that: they thought they needed 10,000 gold examples, then 1,000, then found 100 working pretty well, and he now tells people to get to 10 and start. The internet has inputs and outputs. The thinking in between was never written down.

Nathan is trying to fine-tune a model that writes as he does, and the research on LoRA-style fine-tuning keeps splitting the problem in two. Style, he can teach. Facts, he cannot: the model will not reliably learn that his name is Nathan Labenz. So the facts have to come in as context, which sends him into five years of Gmail and Slack — tens of millions of tokens — that he is willing to trust to OpenAI, Anthropic, or Google and not to a random startup he found last week. He is compressing that history toward something like 1 percent of the original and still calling it an unsolved problem for the whole industry. Agents need that idiosyncratic context. Almost nobody has a safe pipe for it.

That, Andrei says, is exactly the vision they are working toward, which is why they have not released the connectors yet. A Google integration is in progress: calendar, Drive, eventually Gmail, so you can chat across those sources the way HubSpot thinks about relationships. Short term, the practical answers are enterprise agreements and the vector stores the big labs already offer, plus a proxy so a power user can point Agent.ai at their own OpenAI-compatible URL — a box under the desk, a private data center, embeddings for millions of financial documents run locally, requests forwarded rather than ingested. Real-time indexing, expirations, and domain-specific retrieval for pharma-scale corpora get complicated quickly. For most white-collar work, he says, a recommended chunk size, an overlap, a retrieval count, and a handful of connectors are most of what you need.

The longer bet is homomorphic encryption on vector stores: encrypt the vectors, query the ciphertext, and decrypt results only with the user's private key, so the platform can do math on data it cannot read. Nobody has shipped that as a commercial, enterprise-grade product yet, which is why it is hard, but he thinks the pieces are there. Apple's Private Cloud Compute, from patents and papers, is the consumer version of the same idea — keep the corpus in an on-device enclave, send only encrypted slivers, return answers only the user can see. Ordinary embeddings are too easy to invert, Nathan notes, which is why a scheme that preserves semantic work without being reversible would unlock a lot of context that currently stays offline. Andrei's same-day advice is older than any of that cryptography: make sure you are paying. If you are not paying $20 or $30 or $100 for API access that comes with privacy guarantees, you are the product, and a free startup will monetize what you give it. He expects key-based approaches to be industry standard in every vector store within six months to a year.

Agent.ai is in the awkward middle Nathan flags: everything is free, at least up to a point, because Dharmesh Shah's backing means they will not go personally bankrupt if the product goes viral. Credits exist, one per agent run for now, and if you burn them you just get more. Andrei knows that is not how costs work once users start stuffing tens of thousands of tokens of context into a run. This year they will likely add a way to buy more tokens, timed with builder-side incentives. Until then the goal is adoption and a map of what actually works. They think they can sustain hundreds of thousands or even millions of users at essentially unlimited usage for some time. That is expensive if you are doing it yourself at 10,000 users. It is one reason he expects consolidation across AI tools, agents, and marketplaces: most have not figured out a business model.

The builder side is the part they want to get right. Incentives might include acquiring or "acqui-hiring" five, ten, or twenty small agent shops to seed the catalog, plus some mix of a flat fee and usage. Users might pay 99 cents a month for something useful, $5, or $20 for unlimited ChatGPT-class usage through the platform with their own data. Credits would rev-share back to the creator, and creators would get tools to charge for a higher tier. They will also keep seeding marketing, sales, and customer-service agents — HubSpot-adjacent personas — and they are not fundraising, which lets them stay focused on product. The intended customer is not the Fortune 500. Later in the conversation Andrei is explicit: they are focused on solopreneurs and companies from zero to about 100 people. If you are an enterprise trying to use Agent.ai, it is not for you.

The same "pick the right tool, do not overcomplicate it, ship" philosophy shows up in the stack. Andrei likes Cursor as much as Nathan does, alongside Copilot and some custom helpers. The backend is Python on AWS, auto-scaled; the frontend is Next.js, React, and TypeScript; the relational store is Postgres, plus some queueing. They ship dozens of times a day. The company is about 12 or 13 people; three or four months earlier it was Andrei, Dharmesh, and a freelancer on the frontend. On vectors they have tried the field. Postgres with pgvector is now as fast as, and on some benchmarks faster than, Qdrant or Pinecone. They still use Pinecone. The debugging and search tools are better out of the box, it is essentially infinitely scalable, and for millions to tens of millions of records it costs dozens of dollars a month. Fifteen milliseconds versus ten does not matter when the wait is a language-model round trip. Instantiating storage, querying, and upserting is about ten lines of code on a connection string. He would rather pay that bill than debug which replica got pegged by a bug in a specific implementation.

Nathan reads the equal billing of Agent Network and Human Network as a polite way of saying the robots are coming for jobs, and he wants the industry to be more honest sooner. He admires Klarna's CEO for saying a customer-service bot was doing the work of 700 people and that headcount in that area had actually come down. Andrei will not take that frame. AI agents, as he looks at them, are highly effective, specialized, semi-autonomous applications, and that is it. They are not coming for anyone's job. They help do certain parts of certain jobs better, faster, and cheaper. A human stays in the loop to decide what good and bad look like, what the selection criteria are, and what broader context the agent should care about. His analogy is the ATM and the bank teller: tellers did not vanish; they moved into customer relationships. He actually thinks people will need to do more jobs, not fewer, just elevated — more sophistication about data and context, more honest conversation about the biases every agent carries, more documentation and constraint so the work gets done faster.

The sentence he puts at the center is that human plus agent, in most of these cases, will win over just agent or just human, and that the work itself should get more enjoyable. Nobody really wants to copy data from one inefficient enterprise system to another all day so they can be creative for thirty minutes. Regulation is the part he will not wave away. Technology is moving faster than any statute, which is a bit scary, and platform builders have to refuse some paid use cases rather than enable every lookalike and voice clone that election season can invent. Guardrails that do not lobotomize the technology, or push the cradle of progress elsewhere, are a hard balance. On the farther social question he does not pretend a startup can answer it. Something like universal basic income, a Star Trek-ish state where money matters less, is probably in the mix, and that requires government intervention rather than private virtue.

Nathan's last scenario is an "agent awakening": model capability has jumped about every two years, everyone is now building scaffolding to cover the weaknesses, and a new release could tip semi-autonomous systems from not-quite-working to working all at once. Should there be speed limits so agents move at human pace? Should they have to identify themselves, perhaps with cryptographic provenance? Andrei throws cold water. We have been almost there for decades. Even if ChatGPT 6 arrived tomorrow, skipped 5, planned, executed, and sat above human intelligence, you would still have to integrate it, change workflows, and define ground truth. The nearer failure mode is that the model gets *better* at making things up, to the point where people cannot tell. Robots cheap enough to plant a brain into, and cars that drive in rain and snow, are further still. The practical drill he offers instead is to list the day's work, ask which 80 percent could be outsourced, and notice which tasks can be structured into a few steps with a clear definition of done. Those are the ones tools can take today. The rest is likely the valuable part.

The private motive sits at the end. Andrei has a three-year-old son, and he is thinking about the world that child will grow into — schoolwork, a first job, talking to something as casually as a kid talks to Alexa. The tools he wants to ship are ones any small-business owner can use and, close enough, build for themselves, without being a Fortune 500. Feedback already includes stories he would not have invented, including people building recipe agents to stay healthier. Human creativity, he says, is kind of unbounded. The team is still about twelve, they are always hiring in go-to-market, marketing, sales, and engineering, and the site is agent.ai. Nathan calls it a hard domain to beat for the opportunity.