{"slug": "how-long-should-an-ai-agent-live", "title": "How Long Should an AI Agent Live?", "summary": "Long-running AI agent sessions degrade performance and pose security risks, so agents should be given a 24-hour life cycle with daily resets and ephemeral sub-agents, according to a new essay by Tomasz Tunguz. The essay recommends a daily coordinator that delegates tasks to single-purpose helpers lasting about 30 seconds, and a nightly consolidation pass that saves durable preferences to a file. This pattern addresses issues like context rot, memory degradation, and security vulnerabilities from persistent access.", "body_md": "Mary, Mary, quite contrary, how long do your agents live?\n\nNew products like [Grok Bot](https://grok.com) & other [meta-harnesses](https://tomtunguz.com/the-harness-is-the-new-battleground/) ask us to create agents. How long should they live?\n\nWhen you design a calendar agent, how long should its session run? Should it stay alive for your entire five-year tenure at a company, or reset every day?\n\nEver more powerful models tempt us to build perpetual sessions that never close. But long-running sessions rot from the inside out.\n\nAs conversational turns pile up, attention degrades. Modern models easily spot a single fact in a long document. But research shows even across frontier models, agent memory is like human memory : it degrades as it grows.[1](#fn:1)\n\nTemporary commands turn into permanent ghosts. Tell your bot in March, “I have a cold this week, cancel morning meetings,” & by November it is still avoiding morning slots.\n\nLong sessions also break security. An agent holding multi-year read & write access to your inbox & calendar is an open door. One malicious email or calendar invite can poison the conversation, quietly hijacking your schedule months down the road.[2](#fn:2)\n\nThe winning pattern is simple : give your daily assistant a 24-hour life, & delegate individual tasks to narrow specialists.\n\nA daily reset matches how humans actually work. During the day, your assistant remembers immediate context : “I’m running fifteen minutes late,” or “keep two to three free for prep.” At midnight, the active conversation wipes clean so tomorrow starts fresh.\n\nWhen work needs doing, the daily coordinator hands the job to a single-purpose helper : a calendar agent to schedule, an email agent to draft a reply, or a news agent to search the web. Each helper lives for thirty seconds with only the specific tools it needs, does the job, & disappears.[3](#fn:3)\n\nFor the daily coordinator, the system prompt acts as a dispatcher :\n\n```\nYou are Tomasz's daily coordinator.\nYour session lives for 24 hours.\n\nWorkflow:\n- Morning: Load preferences from `preferences.md`\n  & today's calendar.\n- Intraday: Do not execute directly. Delegate\n  to sub-agents (`calendar_bot`, `email_bot`).\n- Night: At midnight, save durable learnings\n  to `preferences.md` and terminate.\n```\n\nFor the calendar helper, the system prompt is a stateless executor :\n\n```\nYou are an ephemeral calendar specialist.\nProcess this single request, call the tool,\nreport the result, & terminate.\n\nRules:\n- Time zone: America/Los_Angeles.\n- Duration: 30 minutes.\n- Hours: 9:00 AM – 6:00 PM.\n- Always check availability first.\n  Never double-book.\n- If full, propose 2 nearest openings & stop.\n\nOutput: Return event title, time, & attendees,\nthen exit.\n```\n\nBefore the day’s session wipes at midnight, a quick consolidation pass runs. An offline summarizer reviews the day, saves lasting preferences (“Tomasz prefers thirty-minute meetings”) into a permanent note on disk, & throws away the rest of the daily chatter.[4](#fn:4)\n\nDoes Grok Bot or your chat assistant perform this sleep cycle automatically?\n\nNot today. Most bots leave threads open forever until you click “+ New Chat” or context compaction silently erases your rules.\n\nThe nursery rhyme asks about a garden : silver bells & cockle shells, all in a row. Things that persist, in an order someone chose. The answer for an agent is the same : throw away the conversation ; keep the rules in a file to keep your agent & its garden healthy.\n\n-\nAmirali Ebrahimzadeh and Seyyed Muhammad Salili,\n\n[“Not All Needles Are Found: How Fact Distribution and Prompting Shape Inference in Long-Context LLMs,”](https://arxiv.org/abs/2601.02023)arXiv:2601.02023, January 2026; Kelly Hong et al.,[“Context rot: How increasing input tokens impacts LLM performance,”](https://research.trychroma.com)Chroma Research, 2025.[↩︎](#fnref:1) -\n[“Sleeper Memory Poisoning in LLM Agents,”](https://arxiv.org/abs/2605.15338)arXiv:2605.15338, May 2026. Demonstrates persistent cross-session memory poisoning attacks in stateful AI assistants.[↩︎](#fnref:2) -\nShiyang Chen,\n\n[“Governance Decay: How Context Compaction Silently Erases Safety Constraints in Long-Horizon LLM Agents,”](https://arxiv.org/abs/2606.22528)arXiv:2606.22528, June 2026. Demonstrates that compaction drops standing rules in 30–59% of episodes.[↩︎](#fnref:3) -\nAnthropic,\n\n[“Dreams: Memory Consolidation,”](https://platform.claude.com/docs/en/managed-agents/dreams)research preview`dreaming-2026-04-21`\n\n, April 2026; and the Letta v2 stateful agent framework (2026).[↩︎](#fnref:4)", "url": "https://wpnews.pro/news/how-long-should-an-ai-agent-live", "canonical_source": "https://www.tomtunguz.com/how-long-should-an-agent-live/", "published_at": "2026-08-24 00:00:00+00:00", "updated_at": "2026-08-25 18:44:47.650020+00:00", "lang": "en", "topics": ["ai-agents", "ai-safety", "ai-research", "ai-products"], "entities": ["Tomasz Tunguz", "Grok Bot", "Chroma Research", "arXiv"], "alternates": {"html": "https://wpnews.pro/news/how-long-should-an-ai-agent-live", "markdown": "https://wpnews.pro/news/how-long-should-an-ai-agent-live.md", "text": "https://wpnews.pro/news/how-long-should-an-ai-agent-live.txt", "jsonld": "https://wpnews.pro/news/how-long-should-an-ai-agent-live.jsonld"}}