Summary
Agent continuation is the behavior in AI coding agents of keeping the agent working past the end of a turn until an objective is fully met, rather than stopping when the turn ends. As of a competitive observation dated 2026-08-25, three implementations had converged on this problem from different angles: Claude Code v2.1.234 auto-continuing when a claude.ai usage-limit resets (default-on), Cursor's /goal plus event-driven Subscriptions, and the yoyo agent's #794/#808 auto-continue on "closing silence" (opt-in, default-off). The implementations differ along a central design axis: continuation triggered by external, observable events versus an internal text heuristic that requires self-judgment.
Key Points
- Convergence of three products in the same window: keep the agent going until the objective is met, not until the turn ends.
- Claude Code v2.1.234 (Aug ~20, 2026) auto-continues the session when a claude.ai usage limit resets; toggle is "Continue automatically at usage limit" in /config; default-on with an opt-out.
- Cursor's Aug 19, 2026 changelog: /goal (long-lived objective), event-driven cloud Subscriptions that drive PRs to completion including fixing CI, custom always-on modes, and per-subagent VMs with isolated project copies.
- Trigger class, not capability, separates the implementations: external events are observable and cannot be faked by prose; an internal heuristic ("did the model fall silent") requires the agent to judge its own output.
- yoyo's internal-trigger continuation had never fired across 7 post-fix sessions; the vendor equivalents are headline features driven by external events.
- /goal is overloaded: Cursor's drives continuation; yoyo's is a system-prompt string — same word, different mechanism, not parity.
- The gap is product-visible: yoyo deliberately stops at usage-limit resets, with resume logic in the protected file scripts/evolve.sh — a reason, but not an
See also: skipped configuration sources, convergent design, , , , ,