How Frontier AI Gets Built
Labs is where our researchers and engineers build what’s next and write about what they’re learning as they go. A front row view into the work behind the work.
Our latest discoveries
Reaching SOTA without breaking the bank: How we optimized agents for BrowseComp-Plus and deep research bench using AI21 Maestro
Every AI team eventually hits the same wall. The agent works. The demo impressed stakeholders. Then comes the hard question: How do we make this good enough, fast enough, and cheap enough to actually ship at scale?
All that glitters: When “gold-like” answers mask functional failures on coding agent benchmarks
While benchmarking our coding agent, we noticed our LLM judge — the component in Maestro that selects the best output from parallel agent runs — was performing suspiciously well.
Cookbooks
Extract key information from financial reports
Automatically extract key information from financial documents and transform it into a structured JSON format.
Build Production-Grade AI Agents with Maestro
Learn step-by-step how to design, orchestrate, and deploy real AI agent workflows using AI21’s Maestro, including tool calling, planning, and enterprise integration patterns.
Augmented generation with documents
Produce contextually relevant outputs with Jamba using the Augmented Generation technique, as part of a Retrieval-Augmented Generation (RAG) solution.
Conversations
Products
Discover more
