The #1 Rule for Building AI Agents in 2026 | Yutori, Abhishek Das
In a Nutshell
Current AI web agents fail on multi-step workflows due to compounding errors, even at 90% per-step accuracy, making first-try reliability non-negotiable—Yutori rejects tolerance for retries or non-determinism. Yutori, co-founded by AI researchers like Abhishek Das, builds trustworthy agents for web tasks via rigorous evals, self-correction, backtracking, and proof-of-work transparency (e.g., Scouts product shows agent steps). Success demands 80/20 feature prioritization blending user feedback with intuition, team dogfooding, and obsessive craft to create magical, reliable products.
These notes were generated by AI and may contain inaccuracies.
There are about 100 different agent products claiming to do anything on the web, but they often fail on the first try. For a 10-step, 20-step, or 50-step workflow, even 90% accuracy per step leads to rapid compounding of the 10% error rate, resulting in very low overall success rates. This is why technology isn't ready for long-horizon workflows. We've started normalizing tolerance for non-determinism and low reliability in products, but this is unacceptable for agentic products. > If it's not good enough to work on the first try, it's not good enough.
Abhishek Das is co-founder and co-CEO of Yutori, building agents that take actions and complete tasks on users' behalf on the web, freeing users to focus on meaningful work. The three co-founders are all AI researchers by background. Investors Faith Lee and Jeff supported them. Abhishek comes from a family of doctors and medical practitioners but is scared of blood, so he pursued engineering instead.
Science and the scientific process appeal to him: hypothesizing, designing experiments to validate/invalidate, drawing conclusions, and iterating. He attended IIT Roorkee for undergrad, meeting some of the smartest people he knows. In first year, he got good grades in electrical engineering (geared toward power systems), but lost interest and rebelled by stopping heavy focus on it, marking a significant life fork.
Instead, he spent time learning programming and building software. IIIT Roorkee (13-14 years ago) had a strong programming club culture, especially SDS Labs—a group of 10-15 coders from all years tinkering, hacking, and building internet applications for campus. Seeing user interactions provided a dopamine hit, fueling late-night feature additions and improvements, especially surrounded by obsessed peers.
Abhishek always wanted to start something; considered it after undergrad and PhD but didn't for various reasons. Lots of interesting problems to solve, preferring to push his own vision. Over the last 2-3 decades, web browsers have stayed the same: open, click, scroll, type. Now, opportunity to reimagine with AI assistants taking actions and completing tasks. Future involves talking to AI assistants, with agents working proactively in the background. Digital agents will precede physical agents, with shorter timelines.
In 5-10 years, web interaction will be at a higher abstraction level. Delegate digital chores to AI agents to focus on meaningful tasks, enabling humans and agents to collaborate for productivity gains, not replacement. Makes web accessible, e.g., parents no longer need to learn every new website—just instruct the assistant.
Many agent products claim broad capabilities but fail initially, with a normalized view that "it usually works" after 10 tries (maybe 3-5 successes). Agents make sequential decisions; error compounding in multi-step workflows explains low success. Technology not ready for long-horizon tasks. Key: recognize mistakes, backtrack, and explore different branches.
Yutori invests in evals and guardrails: every production query runs through comprehensive evals to identify strengths, weaknesses, and domains needing work. For web agents, can't train on every website (new ones emerge constantly, vast numbers exist). Models will err on new sites like humans do (wrong buttons). Crucial: recognize mistakes, backtrack, and correct.
Ecosystem tolerates non-determinism, slop, and poor reliability in agentic products. > If it's not good enough to work on the first try, it's not good enough.
Take an 80/20 approach: prioritize top 10 of 100 features, informed by users but also intuition for unrequested features that delight. Example: iOS/Android auto-fills 2FA SMS codes—saves seconds daily worldwide, feels thoughtful for tiny paper cuts, even if unasked for. Marriage of user feedback and intuition.
In era of easy prototypes with coding LLMs, true differentiator is taste and craft in intuitive, well-designed products. Team dogfoods seriously: weekly 1-1.5 hours; run tens of internal experiments, shipping ~1 to production. Builds taste for what feels awesome/magical through reps.
Grad-CAM led by labmate, Abhishek supporting author at age 25. Extremely well-received (20-30,000 citations). Focused on interpretability in deep learning classification models: what image parts do models use for predictions (thousands of categories)? Combines image signals and label signals for intuition. Works effectively across tasks/models to this day.
AI models must convey not just final prediction/answer but proof of work—steps taken. Grad-CAM exemplifies this. In Yutori's Scouts product, set up agents to monitor web, generate reports, notify on value. UI button inspects behind-the-scenes: sites visited, info pulled—builds trust.
Yutori focuses on getting it right: reliable, trustworthy products. Attention to visible details builds trust in invisible parts. Approach everything with hard work and detail, like awesome products around us. Takes time to build meaningful, delightful, reliable experiences bringing future visions to life.
Building delightful and reliable product experiences doesn't appear out of nowhere.
Keep EO in your library
Save the videos and channels worth coming back to, and find them again in one place.
