Automation and Background Work
Definition
Automation and background work are the harness capabilities that let agents execute tasks outside the narrow rhythm of a live chat turn. The common pattern is simple: define a job, let it run later or elsewhere, and return the result to a review surface instead of pretending the user must supervise every keystroke in real time.
Representative models
codex-cli now participates in this through Codex app automations, cloud-delegated background tasks, and automatic GitHub review. claude-code documents recurring tasks, CI usage, and remote-control flows as part of the main product story rather than as side experiments; its scheduling docs also distinguish durable cloud or desktop tasks from session-scoped /loop jobs. hermes-agent treats cron-style natural-language automations as a core feature of the runtime.
What background work requires
- Durable state so the job can start without relying on a warm transcript.
- Clear work objects and destinations, which is why this topic overlaps with work-management-primitives.
- Explicit execution locality: some jobs run in a fresh cloud clone, some on a local machine, and some only inside a live session.
- A review or inbox surface where results can be inspected before the system quietly compounds its own mistakes.
- Permission boundaries suitable for unattended execution, because an agent running later is still an agent with tools now.
Design lesson
Background execution turns a harness from an interactive helper into an operating system for delegated work. That is powerful, but it also means automation quality depends on context-engineering, evaluation-and-review-loops, safety-and-permissions, and the chosen orchestration-topologies rather than on model cleverness alone.
Related pages
Read this with work-management-primitives, evaluation-and-review-loops, orchestration-topologies, codex-cli, claude-code, and hermes-agent.