We built background agents that automate recurring work (in production). Unlike interactive agents where users can fix mistakes, background agents run when no one is watching.
Here's a bunch of approaches that we found useful: making everything a tool call (including stopping), treating tool metadata as prompts and rewriting them all, using deterministic binary checks with cheap LLMs plus optional review from expensive LLMs for quality control, and using multiple simple memory approaches instead of one complex system.
The patterns might be useful if you're building agents that need to run reliably without supervision.
We built background agents that automate recurring work (in production). Unlike interactive agents where users can fix mistakes, background agents run when no one is watching.
Here's a bunch of approaches that we found useful: making everything a tool call (including stopping), treating tool metadata as prompts and rewriting them all, using deterministic binary checks with cheap LLMs plus optional review from expensive LLMs for quality control, and using multiple simple memory approaches instead of one complex system.
The patterns might be useful if you're building agents that need to run reliably without supervision.