The frontier problems
in agentic AI.
We research what breaks between demo and production — coordination, memory, browser automation, continual learning, and how you measure whether an agent has earned its autonomy.
Rahul Chinthala · Senior SDE at AWS · Amazon Q infrastructure
Five problems we work on
in production.
Multiplayer AI
GridWhat happens when agents need to coordinate? Shared intent, conflict detection, and the hard problem of making networks of agents coherent — not just individually capable.
02Memory
Agents that forget everything between sessions are perpetual first-days. We build typed memory systems that accumulate context, trace recall to source, and forget explicitly.
03Digital AI Agents
Digital EmployeeThe web is hostile — DOMs drift, sessions expire, modals interrupt. We've kept 27 production automations alive through months of DOM changes without manual intervention.
04Continual Learning
Learning at the system level, not the model level. Typed memory persistence, pattern extraction, skill improvement across runs — agents that get better in their specific context.
05Agency Evaluation
Benchmarks measure outputs. Agency evaluation measures what an agent is allowed to do next — and how that scope expands as trust is built through a track record.
Research made tangible —
systems running in production.
Grid
Coordination layer for multi-agent AI systems
When agents work in parallel, locally correct decisions become globally incoherent. Grid provides shared intent, conflict detection, and mandate enforcement — making networks of agents coherent, not just capable.
Digital Employee
AI agent that learns your organization
27 production browser automation skills across X, LinkedIn, and the broader web. 10K+ typed memory objects. The same agent that executes tasks in the browser responds in Discord with the same context — and doesn't break when the web changes.
What we've learned
shipping real systems.
I Built AI Agent Memory Wrong 3 Times. Here's What Finally Worked.
What seven research papers, one catastrophic bug, and 10,000+ memory objects taught me about building memory systems that actually earn user trust.
The Selector Ladder: Why Your Browser Automation Breaks Every Week
After running 27 browser automation skills across X and LinkedIn for months, we developed a reliability hierarchy that survives CSS obfuscation, DOM reshuffles, and React re-renders.
The Coordination Problem: Why Multi-Agent AI Systems Break at Scale
Single agents are getting powerful. Multi-agent systems are still a mess. The problem isn't intelligence — it's coordination. Here's what we learned building Grid.
Agents That Actually Learn: Beyond Session-by-Session Amnesia
Every session starting from scratch is not a feature — it's a ceiling. Here's how we think about continual learning for production AI agents.
Senior SDE at AWS, where he built Amazon Q's cost analysis infrastructure and the AWS BCM MCP Server. Started Ollie Labs to research the reliability problems that sit between a demo and production — how agents remember across sessions, coordinate without colliding, navigate a hostile web without breaking, and how you know when to trust them.
The lab is small by design. Every finding comes from systems that are running.
Working on the same problems?
If you are building multi-agent systems, running browser automation at scale, or thinking about how agents earn trust — we want to compare notes.
rahul@ollie-labs.com