# Drift over time

Long runs lose the thread.

**Is a:** [Failure kind](/reference/ontology/unattended-agent-failures/failure-kind)

**Guarded by:** [Fresh starts](/reference/ontology/unattended-agent-failures/fresh-starts)

**Grows with:** [Compounding failure](/reference/ontology/unattended-agent-failures/compounding-failure)

Long runs lose the thread. Cursor says that after running agents for long stretches "we still need periodic fresh starts to combat drift and tunnel vision" ([Cursor, Scaling long-running autonomous coding](https://cursor.com/blog/scaling-agents)). Software that stays in service ages, and a planned restart was found cost-effective at AT&T Bell Labs ([Wikipedia, Software aging](https://en.wikipedia.org/wiki/Software_aging)).

Guard: [fresh starts](/reference/ontology/unattended-agent-failures/fresh-starts). It grows with run length, as does [compounding failure](/reference/ontology/unattended-agent-failures/compounding-failure).

---

Published 2026-10-29. Web page: https://www.mackenziebowes.com/reference/ontology/unattended-agent-failures/drift-over-time