Introducing Infinity: a runtime for extreme agent concurrency

This article introduces Infinity, an open-source agent runtime that reimagines how agents execute by suspending them during tool calls, enabling extreme concurrency and serverless deployment.

• Agents suspend when calling tools, saving conversation history and resuming when results arrive, similar to stackless coroutines in programming languages
• Supports 75,000 concurrent agents on 8 GB of memory (Raspberry Pi 5) with ~100 KB per idle agent
• Runs on serverless platforms like AWS Lambda with SQS and Aurora DSQL, requiring no compute while agents wait
• Uses synthetic tool calls to deliver asynchronous events within model API constraints
• Enables agent-to-agent communication through spawn_thread and report_to_parent built-in tools
• Processes events in child threads by default to avoid context window bloat in parent agents

Infinity enables long-running, event-driven agents to run efficiently on both local processes and serverless platforms by treating tool calls as suspension points.