Background Agents

The default way to work with a coding agent is a conversation. You write a prompt, watch tokens stream, correct when it drifts, and prompt again. It feels productive because it is fast, but it has a hard ceiling: it spends one human's full attention per agent. The agent is quick; the supervising human is the bottleneck.

The shift underway in 2026 is the move from this synchronous, supervised loop to asynchronous, fire-and-forget work. You dispatch a task to a cloud agent and walk away. It runs in an isolated environment, survives across sessions, and notifies you when it finishes or when it genuinely needs a decision. Scheduled routines run on their own cadence with no trigger from you. Fleets of agents run in parallel on independent slices of a larger job.

The leverage is where the value is. But it is also where most teams get hurt, because the property that makes async valuable — nobody is watching — is the same property that makes it dangerous.

This whitepaper covers the patterns Tenten AI uses to run unattended agents in production: the infrastructure beneath them (durable execution, queues, sandboxes, notifications), the verification and observability that make "nobody is watching" safe, and the discipline of writing tasks an agent can complete and self-check without a human in the loop.