E45: The AI Copilot Revolution with Div Garg of MULTI·ON

Episode Summary

Executive Summary: Div Gerg explains MultiOn’s browser-based AI agent vision: make delegation as easy as giving natural-language goals, then let agents execute web tasks while users supervise. The conversation covers product strategy, safety, memory, skills, reliability, and how agents may turn humans into coordinators of parallel AI work rather than single-threaded operators.

Main Topics: MultiOn’s product strategy and browser-extension approach (Priority: 5/5): Div argues that embedding the agent in Chrome is the best onboarding and interaction model because it uses existing logins, allows supervision in real time, and reduces friction compared with cloud-only or OS-level systems. Agent architecture: planning, representations, and action grammar (Priority: 5/5): The system combines DOM/image representations, a planner model, and an intermediate action language that can be type-checked before execution. Div says this improves generalization, safety, and reliability versus direct code generation. Reliability, task success, and critic loops (Priority: 5/5): A major challenge is reducing variance and ensuring the agent understands when a task is truly complete. MultiOn uses success prediction and experiments with critic models and user confirmation steps to validate outcomes. Skills and memory as high-level natural language abstractions (Priority: 4/5): MultiOn’s skills are intentionally high-level, letting users define reusable workflows in natural language that can be compiled into runtime actions. Memory is split between explicit profile data and learned personalization. Safety, prompt injection, and guarded rollout (Priority: 5/5): Div emphasizes closed beta, blacklisted sensitive sites, type-checked action code, protected skills, and a kill-switch mindset. He sees prompt injection and unauthorized actions as the key security risks. Timeline for adoption and the future of work (Priority: 4/5): Div believes simple tasks are nearly ready for non-technical users within months, but more complex multi-step delegation still needs better foundation models. He sees agents as supplements first, eventually becoming coordinators of AI labor. Research-to-product feedback loop (Priority: 4/5): The company aims to gather interaction data from real users, fine-tune models, and improve the product through usage. Div views this as a major advantage over teams that stop at demos or research papers.

Key Arguments: Browser-based agents are the right starting point because they inherit existing logins and reduce authentication friction, making delegation practical for everyday users. A robust agent should not modify its own source code; preserving fixed boundaries is a core safety tenant to prevent self-evolution or uncontrolled behavior. The system is built around a semantic action grammar rather than raw code, allowing verification, type-checking, and safer execution. Reliability matters more than raw capability for adoption: the main problem is not whether an agent can act, but whether it can act predictably and complete the right task. High-level skills are better than low-level scripts because users can author and update them without technical expertise, and they can be recompiled as websites change. A critic/validator layer can catch mistakes, confirm task completion, and provide correction signals for future improvement. MultiOn is positioning itself first as a supplement to human assistants, with the longer-term hope that agents will increasingly coordinate rather than merely assist. The biggest remaining bottleneck for more complex delegation is model reasoning quality; Div does not believe current GPT-4-level systems can yet handle all long-horizon tasks reliably.

Data Points: Company stage: Closed beta - MultiOn is deliberately limiting access while it safety-tests the product and gathers data. Seed round status: In the middle of a seed raise - Div says announcements should come in a few weeks. Planned launch timing: Stanford launch in the next month - He describes an initial phased rollout before a wider release. Simple-task readiness timeline: Next three months - Div says non-technical users should be able to use the product for simple tasks within this window. Interaction data collected: 50,000 web interaction samples - Used to train/fine-tune in-house models for agent performance. Model type under fine-tuning: Falcon 40B - Div says the team is starting to fine-tune this model. Research cadence: 20 foundation models dropping in a week - He uses this to illustrate the speed of AI progress. Research output cadence: 100 papers published in a week - Another example of how fast the field is moving. User supervision horizon: First few uses - Div says users may watch and intervene initially, then allow background execution later. Planning complexity threshold: 20+ steps - He says current models still struggle with very long, abstract reasoning chains.

Pivotal Quotes: "We want to unlock parallelism for humanity where like currently you can imagine like we are all single-threaded." — Div Gerg: Describing the long-term vision for agents as a way to multiply human productivity through concurrent AI delegation. "An agent should not be able to modify its own source code because if it's able to modify its source code then could like self-evolve it could do like really Weird things." — Div Gerg: Explaining one of his core safety principles for building intelligent agents. "I don't think GPT-4 is there yet when it comes to a lot of really complex reasoning." — Div Gerg: On the limits of current foundation models for long-horizon agent tasks.

Implications: MultiOn reflects a broader shift from chatbots to supervised action systems. If browser-native agents become reliable, they could automate routine web work, reshape assistant roles, and create a new market for safe, reusable AI workflows.