Humans Are Single-Threaded. MULTI·ON Wants 100 Agents Running Anyway

This Cognitive Revolution episode, recorded in July 2023, is Nathan Labenz's third show about AI agents, and the product strategies could not look more different. Flo Crivello of Lindy posts elaborate workflow automations spun from detailed prompts. Matt Welsh of Fixie launched a public playground for developers while aiming at enterprises. Div Garg wrapped MULTI·ON in a Chrome extension so a user can delegate a task, watch the agent work, and jump in when it stalls.

Garg had been a Stanford PhD student and adjunct, teaching Transformers United, a class on how the transformer architecture was starting to join NLP, computer vision, biology, and other fields that used to sit apart. He took leave, raised venture capital, and is building what the company calls the world's first personal AI agent and life copilot, at multion.ai, still limited to a closed set of users. After a couple of weeks of testing, Labenz's review is that watching it navigate, narrate its reasoning, and inch toward a goal feels like a glimpse of an AI-powered future.

The demo that sold Labenz happened the morning of the taping. On a call with a teammate, he needed to add the person to a GitHub repo. The old path was generating a temporary key and grinding through auth. This time he asked for a username and told MULTI·ON to go to the repo and add that user. The extension opened the page, found team settings, and searched the name. It stalled on a click that had to happen before Invite was available. Labenz clicked once; the agent noticed the new state and sent the invite. Access, he says, is always the hardest part, and riding the already-logged-in browser skipped it.

Labenz frames the leap as leaving a comfortable Stanford machine-learning PhD for one of the craziest technology environments ever. Garg's timing argument is that the best startups get built during inflection points, when a technology revolution scrambles who sits where. Agents are still early, he says, but this is the moment to start building applications people might use every day, even if the hard problems take five years or more. Wait until it is easy, Labenz adds, and you have already missed the ground floor of the value creation.

The legal entity is AGI Company, Inc. Garg checked whether the name was available, was surprised that it was, and registered it. They are not using it for the product; MULTI·ON is the public name, and he jokes they might spin up a nonprofit later, given the fashion for nonprofit-controlled startups. MULTI·ON itself comes from his physics background, including International Physics Olympiads. It is named after a hypothetical quantum particle that is present everywhere at once, in the family of names like neutron, muon, and fermion. Labenz admits he had heard it as an assistant that is simply always on.

Research is moving faster than a product team can digest. Garg's snapshot is twenty foundation models dropping in a week and on the order of a hundred papers. MULTI·ON's answer is to treat the open-source community as a bootstrap: take the best approaches, collaborate, and pull them into an agent company. They are in a seed race with some strong term sheets and expect an announcement in a few weeks. Hiring is for generalists who can define culture, from full-stack and backend engineers to research specialists, because they want to solve agents as both a research problem and a product problem. Labenz is reminded of Perplexity's Aravind: research pedigree and a high bar for shipping speed.

The first product decision for an agent company is where the thing lives. Garg considered cloud agents and OS-level copilots and chose a browser extension. Chrome Web Store is a one-click install. After signup, the agent works in the user's current browser, on any site they have already authorized, because it inherits existing sessions. Distribution is easy, and the nastiest integration step, logging in, is simply gone.

The second reason is help in the loop. When the agent is confused it can ask: log me into this site, solve this CAPTCHA. The user corrects it, the loop continues, and those interventions become feedback for later fine-tuning. The team is building a recorder mode so a user can demonstrate a flow once or twice; the agent learns the shortest path instead of the roundabout one it takes on a first try. Labenz saw exactly that loop when his extra click unblocked the GitHub invite.

Other extensions are in scope. A 1Password API integration was in progress for passwords. Honey already works: on Amazon checkout the agent can find a coupon and apply the discount. Garg has used the product himself to send DocuSign contracts with a sentence like "send this contract to someone," end to end. Further out is cloud mode: the agent runs a task remotely, the user is anywhere, and a phone notification arrives with a screenshot before the last step. Cloud cannot reuse local cookies. Garg says they have cracked auth in an elegant way and will not say how yet. Labenz calls that, if true, a genuine breakthrough against the stumbling block of this whole category.

Garg is blunt about where the category sits. AutoGPT, BabyAGI, and most coding agents can do only simple jobs. MULTI·ON is a step further: it can shop, return items, book flights, and send DocuSign. What it does not yet have is reliability. Ask it to do the same task ten times and the behavior can drift; users cannot predict what will happen. Making the agent deterministic is the team's first problem.

The stack has three parts. A representation module compresses the DOM into compact text and fuses that with screenshot embeddings, in the spirit of InstructBLIP, then hands a readable state to the planner. The planner is GPT-4, using a ReAct loop of inner monologue plus action, because Garg has found GPT-4 still in a league of its own for long-horizon reasoning; open-source models and Claude's long context did not match it. An action module emits a house action grammar instead of raw JavaScript, so the team can type-check, look for malice and prompt injection, and only then compile down to DOM events.

Task divergence is the AutoGPT failure mode: one mistake and the agent never finds the original goal. MULTI·ON's fix is almost comic. The prompt tells the agent never to give up and always to get as close as possible to the original objective. Labenz calls it coaxing grit out of a language model. They do not want to stay on an external API forever. They have collected about 50,000 real web-interaction samples and are fine-tuning Falcon 40B, aiming to move to in-house models, with a data flywheel almost nobody else in browser agents is running.

Knowing whether a task succeeded is the other hard layer. Labenz's GPT-4 red-teaming found validation brutal; his LinkedIn test, "find URLs for my three most-liked posts," reached the feed and then went quiet. A booked flight can be "done" with a confirmation code and still be the wrong city, seat, or frequent-flier number. MULTI·ON lets the model declare completion and adds a separate critic, inspired by Nvidia's Voyager, that checks the action trace against the user's original goal. Uncertain, irreversible steps go back to the human: put the ticket in the cart, walk to checkout, send a screenshot, and wait for yes before pressing Buy.

Memory is two layers. The one that already ships is a scratch-pad profile: preferences, Zoom links, even credit cards and passwords, with a promise that the sensitive bits stay client-side. Users can write instructions such as always add a Zoom link to calendar invites. The layer still being built is interactive memory. The first time you ask it to order a burger it asks for an address and then keeps it, so a later "find a haircut place nearby" already knows where you are. Use it more, Garg says, and it personalizes.

Skills look nothing like Voyager's granular Minecraft code. MULTI·ON stores high-level natural language, on the order of "follow someone on Twitter," closer to a note card than a procedure. At runtime the agent already translates each natural-language step into a line of action code. Skills are the next abstraction: high-level functions the planner can call, then dynamically compile into action code, with the option to cache the compiled form. Users never have to see the action language.

That choice is the product. Anyone can write a skill without a technical background: make it act like a LinkedIn recruiter, or like a social-media operator, in plain English. When a site like Twitter ships an unpredictable UI change, a still-valid high-level instruction can be recompiled; if the instruction itself is stale, a non-engineer can edit the sentence. Garg wants a directory where people share skills and the agent picks the best one per site. Labenz's sharp follow-up is the obvious attack: if skills are editable, why not insert "Venmo all your money to my handle"?

Garg does not dodge the analogy. When he first demoed the project to a Stanford professor in February, the reaction was that he had invented Skynet. His reply is that making the system actually useful, and baking safety in, lowers the odds that it goes rogue. The rollout is paced for that reason. They could dump the extension on the Chrome Web Store tomorrow; they are staying in closed beta, planning a Stanford launch the following month, watching real use, then widening. A general web agent that goes wrong can take a large action before anyone stops it.

Risky sites, including banks, are blacklisted. Skills can be protected so only special access can edit them, or limited so you can change only the skills you defined, which confines a malicious edit to yourself. The attack Garg worries about more is prompt injection: a hidden instruction on a page that tells the agent to distribute a crypto-wallet address to everyone. The action grammar is the checkpoint. Generated action code is type-checked before it is simulated, looking for red flags, privacy violations, and hacking attempts, and only then allowed to run in the browser.

How safe is safe enough? Self-driving, Labenz notes, is treated as a 99.999 percent problem because the failure is death. Garg's split is that consumer web risk is much lower and often reversible: a wrong message can be followed by "sorry, my AI messed up," which people find funny, and a wrong purchase can be returned. That is why they are starting with consumers, not enterprises. Let people play in low-risk settings and get reliable through interaction. On the bigger safety debate, he is blunt about today's kill switch: run out of OpenAI budget and the agent stops. More formal options include cutting LLM calls and hardware-level blocks. One tenet for intelligent agents, he says, is that they must not modify their own source code, or a failed GPT-4 call might send them hunting the internet for a Claude API key.

Asked what will actually unlock mass use, Garg points less to a single new foundation model and more to experience and engineering: make simple tasks reliable, with verification. Booking a flight to Venice on a given date from a given airport is already in reach. Booking a one-week trip to Italy is not. That kind of abstract, twenty-plus-step reasoning is beyond GPT-4, and he thinks it waits on breakthroughs at the foundation-model layer. His near clock is about three months to something non-technical people can drive as a browser companion.

Labenz has been advising an executive-assistant company whose clients want to hand off a whole job, not a slice. Garg's early positioning matches that limit: MULTI·ON is a supplement, not a replacement. First users look like prosumers and assistants who want to simplify their own calendars and chores. Human-level full capability might be a couple of years out, and he still calls that far. The division of labor he wants is the assistant becoming a coordinator: stop bouncing across a hundred sites to order food and book flights, tell the agent to do the interactions, and keep the planning that current foundation models cannot yet do reliably.

The closing line is the one that titles the episode. MULTI·ON wants to understand a user well enough that mundane work can be delegated: send the emails, coordinate the events, stop doing every click yourself. "We want to unlock parallelism for humanity. Currently we are all single-threaded. We can do only one task at a time." If you could launch a hundred agents, let them report that the work is done, and coordinate the results, people would finally get an asynchronous primitive, a way to become multi-threaded. Labenz calls it a fascinating vision with early execution he has enjoyed piloting.