The Quest for Autonomous Web Agents with Div Garg, Cofounder and CEO of MultiOn

Episode Summary

Executive Summary: Multion CEO Div Gerg argues that AI agents remain early: language models can appear capable but still struggle with deep reasoning, unfamiliar interfaces, and long, multi-site workflows. Multion is targeting useful short tasks while developing memory, online learning, trajectory-based planning, and agent coordination. Gerg emphasizes speed and efficiency alongside reliability, careful evaluation, privacy, and safety, and predicts agents will automate digital chores while changing—not simply eliminating—the nature of work.

Main Topics: Why AI agents still fail: Gerg attributes unreliable agent behavior to weak logical reasoning, poor grounding in the environment, and limited experience with real workflows. Websites are complex even for people; agents need ways to explore, learn from outcomes, and adapt to unfamiliar interfaces. Multion’s current capabilities and product roadmap: The product is strongest at short, focused tasks on one website, such as finding items and placing them in a cart, sending an NDA, or scheduling a meeting. The roadmap moves from short tasks to longer and combined workflows, then to parallel agents, scheduled tasks, mobile use, and API-based automation. Learning, planning, and agent architecture: Multion is exploring online learning and reinforcement-learning methods, action-oriented and multimodal models, trajectory-level optimization, and hierarchical plans that refine broad goals into executable steps. Gerg also sees potential in coordinating multiple agents once a single agent is reliable. Memory, context, and efficiency: More context can confuse current models, so Multion manages what information reaches them through retrieval and external memory, including model-selected memory operations. Gerg argues that efficiently representing relevant information and filtering noise will remain valuable even as foundation models improve. Evaluation, speed, and operating costs: Because websites change, benchmarks alone are insufficient; Multion combines automated scenarios with manual, end-to-end testing. The company aims for agents to be substantially faster than people while improving task success, and is pursuing smaller models, caching, and reusable skills to lower inference costs. Safety, privacy, and trust: Gerg describes protections against prompt injection and harmful actions, including monitoring and verification before browser actions execute. He says Multion avoids training directly on users’ private authenticated data and argues that moderation checks and safety-minded prompts should be standard for production agent systems. Workforce changes and the future of jobs: Gerg expects agents first to complement people by taking on repetitive digital chores and tasks people dislike. He argues that automation will shift the composition of jobs and may create new work involving operating, coordinating, and teaching agents, rather than simply reducing employment.

Key Arguments: Current models can produce convincing language and sometimes useful code while still lacking dependable logical deduction; errors become especially consequential when an agent must act in a real, changing environment. Agents need experience grounded in their environments. Gerg argues that exploration, feedback, and online learning could help them adapt to websites and improve beyond behavior learned from static training data. Multion should prioritize practical, bounded tasks first. Gerg identifies short, single-site to-dos as its current strength and treats cross-site composition and longer workflows as subsequent challenges. A stronger foundation model will not make systems work automatically. Efficient context representation, memory, task decomposition, and useful action representations can continue to improve both performance and cost. Product usefulness depends on more than completion accuracy: latency and cost matter. Gerg frames the value proposition as performing selected tasks much faster than a human, without sacrificing reliability. Evaluation must include real-world, end-to-end tests and human inspection because dynamic websites can make static benchmarks misleading. Agent developers should plan for abuse before capabilities scale. Multion’s proposed safeguards include prompt-injection detection, action verification, behavioral controls, and minimizing exposure of users’ private data. Automation is likely to remove some repetitive digital chores and change job responsibilities; new roles may involve supervising, coordinating, and improving agents.

Data Points: Simple-task length: Fewer than 20 steps - Gerg’s estimate of the number of agent steps a simple task may require. Information-gathering task length: About 100 steps - Gerg’s illustrative estimate for a research or information-gathering task. Cost per step: Up to about $0.10; roughly $0.02–$0.03 with more efficient models - Gerg’s approximate current cost figures, distinguishing a higher-cost setup from more efficient models. Estimated cost for a 100-step task: Potentially under $2–$3 - Gerg’s estimate if more efficient models reduce per-step costs to roughly two or three cents. Average model input size: No more than 5,000 tokens - Gerg’s estimate of the average prompt size at an inference step in Multion. Target speed relative to human performance: At least 10× faster - Gerg’s stated product-value benchmark for tasks agents can perform at a comparable level. Reported reach of the demonstration tweets: About 5,000 people - The host said a Multion-generated tweet reached this many people without commenters questioning its origin. GPT-4 Turbo context window mentioned: 128,000 tokens - The host cited this as an available context length while discussing concerns that models may lose focus with very large contexts.

Pivotal Quotes: "the deep work and like the deep logical connections is still not there." — Div Gerg: His explanation of why fluent model output can overstate the actual reasoning and reliability of current AI systems. "we want it to be a product, not research." — Div Gerg: His rationale for optimizing agents for speed and usability as well as task performance. "those shitty jobs won't exist because technology will just solve the problems better." — Div Gerg: His argument that agents will automate undesirable digital chores and alter the nature of work.

Implications: Near-term agent adoption is most plausible for bounded, repetitive tasks, not unrestricted digital delegation. Reliability, speed, cost, privacy, and safeguards will determine whether agents earn trust. As capabilities grow, businesses should expect workflows—and the skills workers need—to shift toward supervising and coordinating automation.