Varick Agents #1 - Context Windows Change Everything

Published about 2 months ago • 1 min read

Thanks for subscribing to the Varick Agents newsletter - there are now over 1,000 of you! I'm Vasuman, and along with my co-founder Eyad, we build AI agents that have already saved our clients over $250,000/mo. We're both software engineers at big tech companies, and we're here to share the real techniques that separate working agents from broken ones.


For the first volume of the newsletter, we're giving you some information we haven't covered on Twitter yet (which you should definitely follow if you don't already). As an added bonus, Vasuman just dropped his first YouTube video breaking down exactly how we built AI Agents that save businesses a quarter million dollars every month.


Context Windows Change Everything


Most AI agents fail because they're built stateless - they forget everything between interactions. But here's what we've learned building production agents: the magic happens when your agent remembers.


Stateless Agent:


  • Each interaction starts fresh

  • Agent asks the same qualifying questions repeatedly

  • No learning from previous conversations

  • Example: A customer service agent that asks "What's your account number?" every single time


Stateful Agent:


  • Maintains conversation history and learned preferences

  • Builds context over multiple interactions

  • Remembers what worked and what didn't

  • Example: The same customer service agent that says "Hi John, I see you're calling about your premium account ending in 4829. Are you still having the billing issue we discussed yesterday?"


When to Use Each:

Go Stateless When:

  • One-off tasks (data processing, content generation)

  • High-volume, low-complexity interactions

  • Privacy-sensitive applications

Go Stateful When:

  • Multi-step workflows spanning days/weeks

  • Customer-facing interactions

  • Complex problem-solving that requires iteration

The stateful approach is why one of our e-commerce clients' support agent now resolves 73% of issues on first contact, compared to 31% with their old stateless bot.

Pro Tip: Store context in structured JSON with timestamps. Most developers dump everything into a string and wonder why their agent gets confused. The streets refer to these as Context Profiles, and we'll deep dive into them in our next newsletter.

What's Next

Vasuman just published his first YouTube video covering how our agents generate massive savings for businesses. Think of it like an intro to the channel but with some sauce, and future videos will dive deep into technical implementations like today's context strategy.

If you're ready to build an AI Agent for your business, we offer discovery calls where we audit your current automation and show you exactly where context-aware agents could save you money. Get in touch with us here.

Talk soon,

Vasuman & Eyad