Why AI agents still fail at multi-step tasks — and what to do about it

Executive overview

Most AI agent products ship without working reliably. Error rates compound across long workflows: at 90% accuracy per step, a 10-step task has only a 35% success rate. The industry has quietly normalised this.

Yutori's approach is to treat reliability as non-negotiable from day one, combining comprehensive evals, self-correction, and product craft to build agents users can trust.

Unreliability in agentic products is a choice — one the best builders should refuse to make.

Why agents break at scale

• A 10-step workflow at 90% per-step accuracy has an overall success rate well below 50%
• Error compounds fast — 20- or 50-step workflows are nearly guaranteed to fail at that rate
• Most products paper over this with "it usually works if you try several times"
• That normalisation of non-determinism is the core problem, not just a known limitation

What reliability actually requires

• Agents must recognise when they make a mistake and backtrack — not just push through
• Every production query runs through a comprehensive eval suite to flag weak domains
• New websites always exist outside training data; robust error recovery matters more than memorised paths
• Guardrails are built into the model training loop, not bolted on afterward

Product craft as a differentiator

• In a world where anyone can prototype fast with LLMs, taste and craft separate durable products
• The team dog-foods new features for 90 minutes every week; tens of experiments run internally before any ships externally
• Small, unasked-for features — like auto-filling 2FA codes — make users feel seen
• User requests inform priorities, but intuition drives the features users didn't know to ask for

Transparency as trust

• Users can inspect every Scout report to see which sites were visited and what the agent looked at
• This "proof of work" visibility is directly descended from Grad-CAM: showing what the model attended to, not just the output
• Attention to visible detail signals reliability in the invisible parts of the product
• Trust is built incrementally; it cannot be declared

The longer arc

• Digital agents will arrive before physical agents — the timeline is shorter
• The future interface is a higher level of abstraction: tell an assistant what you want, not how to click through a site
• The goal is humans and agents working together for productivity, not replacement
• Accessibility is a real benefit: non-technical users no longer need to learn every new website