CHAPTERS
- 0:00 – 0:31
Why most web agents fail: compounding errors and “normalized” unreliability
The conversation opens with a critique of today’s agent products that promise broad capability but break on first use. Abhishek explains why long-horizon tasks collapse under compounding step-by-step error, and argues the industry is becoming dangerously tolerant of non-determinism.
- •Many agent demos don’t work reliably on first try despite big promises
- •Long workflows amplify small per-step error rates into low end-to-end success
- •The ecosystem is normalizing “usually works” quality for agentic products
- •For real adoption, agents must meet higher reliability expectations
- 0:31 – 1:01
Yutori’s mission and founding context: serious agents, not just another demo
Abhishek introduces himself and Yutori’s goal: building agents that take real actions on users’ behalf on the web. He frames the company as a larger bet, backed by respected figures, and grounded in an AI research background.
- •Abhishek Das: co-founder and co-CEO of Yutori
- •Yutori builds web agents that complete tasks for users
- •The founding team has AI research roots
- •Positioned as a bigger vision with notable early support
- 1:01 – 1:31
From a medical family to engineering: attraction to the scientific method
He shares personal background—growing up around medicine but choosing a different path. What pulled him toward engineering was a love for hypothesis-driven thinking and the scientific process.
- •Family background in medicine, but he chose not to pursue it
- •Motivated by fear of blood and a pull toward engineering
- •Strong affinity for the scientific method and experimentation cycle
- •Early mindset shaped by hypothesis → experiment → conclusion loops
- 1:31 – 2:32
IIT Roorkee: pivoting away from electrical engineering into software building
At IIT Roorkee, Abhishek initially did well academically but quickly realized the core EE track didn’t match his interests. He made an early “rebellious” pivot—deprioritizing the curriculum and going deep on programming.
- •Strong early academic performance, followed by an interest mismatch
- •EE focus (e.g., power systems) didn’t align with his goals
- •Decisive pivot at the end of first year
- •Self-directed shift into programming and software development
- 2:32 – 3:32
SDS Labs and the “dopamine hit” of shipping: learning by building for real users
He describes SDS Labs, a small group building campus intranet apps, as formative training in product thinking. Shipping features, watching user behavior, and iterating fast created a powerful feedback loop that fueled his obsession with building.
- •SDS Labs culture: tinkering, hacking, and shipping intranet apps
- •Motivation came from seeing real users interact with what they built
- •Late nights iterating on features and improvements
- •Community of equally obsessed builders accelerated growth
- 3:32 – 4:03
Why start a company: choosing your own problems and pushing a personal vision
Abhishek explains he’d wanted to found something for years, considering it after undergrad and after his PhD. Ultimately, the motivation was to pursue meaningful problems aligned with his own vision rather than executing someone else’s.
- •Longstanding desire to start a company
- •Multiple earlier moments where founding was considered but deferred
- •Drive comes from solving interesting real-world problems
- •Preference for pushing a personal vision over working on others’ visions
- 4:03 – 5:03
The last generation to use a browser: reimagining web interaction with proactive agents
He argues the browser paradigm has barely changed for decades, but agents will shift interaction to a higher-level interface. In this future, users delegate “digital chores” to assistants that act in the background, improving productivity and accessibility.
- •Traditional browsing (click/scroll/type) is ripe for reinvention
- •Future: talk to assistants that take actions on the web
- •Digital agents likely arrive before physical agents
- •Agents + humans collaborate; goal is productivity and meaningful focus
- •Accessibility benefit: reduces the need to learn each new website
- 5:03 – 6:34
Stop normalizing broken agents: long-horizon workflows, backtracking, and guardrails
Returning to present-day reality, Abhishek critiques agents that work only intermittently. He outlines what’s required for robust web agents: recognizing mistakes, backtracking, and systematic evaluation/guardrail infrastructure to understand failures by domain.
- •“Works 3 out of 10 times” is unacceptable for agentic products
- •Agents are sequential decision-makers; errors compound across steps
- •Key capability: detect mistakes, backtrack, and choose alternate paths
- •Heavy investment in evals/guardrails for every production query
- •Use evals to map where agents work vs. where they need improvement
- 6:34 – 7:35
The open web problem: you can’t train on every site, so correction mechanisms matter
He emphasizes a core constraint for web agents: the web is too large and constantly changing to cover exhaustively in training. Because mistakes are inevitable on new sites, the differentiator becomes how well an agent recovers and self-corrects.
- •Impossible to train on every website; new sites appear constantly
- •Models will always face out-of-distribution UIs and edge cases
- •Mistakes (wrong clicks, wrong buttons) are natural—like humans
- •Critical ingredient: recognizing error and correcting via backtracking
- •Sets expectations for how agents must be trained and shipped
- 7:35 – 8:35
The 80/20 product approach: prioritization plus intuition that makes users feel seen
Abhishek describes an 80/20 mindset for deciding what to build amid countless possibilities. Beyond direct user requests, he highlights the role of product intuition—shipping small, thoughtful features that remove daily friction and signal care.
- •Feature prioritization: focus on the most impactful few out of many
- •Balance explicit user feedback with builder intuition
- •Great products include features users didn’t think to request
- •Example: auto-filling 2FA codes from SMS removes repeated friction
- •Small “paper cut” fixes create disproportionate user delight
- 8:35 – 9:06
Taste and craft as the new moat: weekly dogfooding and rigorous internal experimentation
With prototypes becoming easy to generate, he argues differentiation shifts to taste, craft, and design quality. Yutori builds that muscle through weekly dogfooding and a high-volume experimentation pipeline where only the best ideas reach production.
- •Coding LLMs make first prototypes easy; craft becomes the moat
- •“Taste” defined as intuitive, well-designed, magical-feeling UX
- •Weekly dedicated dogfooding session to evaluate new features
- •Run many internal experiments; only a small fraction ships
- •Repetition (“reps”) builds the team’s product judgment
- 9:06 – 12:20
From Grad-CAM to trustworthy agents: showing proof-of-work and building reliability
Abhishek connects his interpretability work (Grad-CAM) to a broader principle: systems should expose how they reached an answer. Yutori applies this by letting users inspect what agents did—sites visited and evidence gathered—because transparency and attention to detail are central to trust and long-term adoption.
- •Grad-CAM: interpretability to reveal what models “look at” for decisions
- •Principle: models should provide proof-of-work, not just outputs
- •Yutori UI includes inspection of behind-the-scenes agent behavior
- •Transparency is critical for trust in agent reliability
- •Reliability requires meticulous shipping standards and attention to detail
