Building the orchestration layer for AI agents — browser control, desktop automation, inter-agent communication, and reliable tool use.
Agents that can actually do things — browse the web, fill out forms, use desktop applications, coordinate with other agents — need infrastructure that doesn't exist yet.
Current agent frameworks are good at chains and RAG but break down when you need an agent to click a button on a website, read the result, decide what to do next, and maybe hand off to a specialized agent for part of the task.
We're building the plumbing: browser automation that works reliably, computer use APIs that feel natural, protocols for agents to talk to each other, and orchestration that doesn't fall over when one agent fails.
What we're building, testing, and learning.
Connecting Claude's computer use capability to Playwright for reliable browser automation. Goal: agents that can navigate real websites without custom per-site code.
Building reusable MCP servers for common tools: file system, databases, APIs, calendar, email. Focus on making them production-ready, not just demos.
Designing how agents pass tasks and context to each other. When should a 'researcher' agent hand off to a 'writer' agent? What context transfers?
How do you give an agent memory that persists across sessions? Exploring embeddings, structured summaries, and explicit knowledge graphs.
Things we're still figuring out.
How do you debug an agent that's autonomous? Traditional logging doesn't capture 'why it decided that'.
What's the right granularity for agent specialization? Lots of narrow agents vs. fewer capable ones?
How do you handle graceful degradation when an agent can't complete its task?
Can agents learn to use new tools from documentation alone?
Interested in this research? Have a related problem?
Let's talk →Reach out to us at info@deepklarity.com