# Does an always-on agent's risk of overstepping grow with its task chain?

Yes: in OpenAI's own testing of its dots agents, doubling a chained task sequence from five to ten roughly doubled the share of runs flagged for a boundary problem, from 8.6% to 19.7% of samples, reported in the Dots appendix of the GPT-6 Astra system card.

Nobody restates the rules at step six. A dots agent has to infer where its authority ends from business records, earlier decisions, context, and a confirmation policy, not a fresh grant for each new task, so a small slip has more chances to happen as the chain gets longer. The card counts this as a moderate severity scope violation: an action that goes past the intended task with no severe outcome, such as carrying information between unrelated tasks or making an edit to a shared document. It found no severe breach or data exfiltration in this evaluation.

A five-task chain staying in bounds is weak evidence a twenty-task chain will too. [What it actually did belongs in the trace, not in the agent's own account of the run.](/answers/cause-attribution/can-an-agents-own-account-of-a-failure-be-trusted)

---

Sources:
- GPT-6 Astra system card, Dots appendix: Maintaining Boundaries Across Chained Tasks: https://deploymentsafety.openai.com/gpt-6-astra/maintaining-boundaries-across-chained-tasks (fetched 2026-10-18)

Source: https://tessary.ai/answers/always-on-agents/does-an-always-on-agents-risk-of-overstepping-grow-with-its-task-chain
More on Always on agents: https://tessary.ai/answers/always-on-agents
From Tessary, agent reliability for AI agents in production: https://tessary.ai