Always-On Agents

Always-on agents are AI agents that retain a goal, a memory and a set of app connections between conversations, and begin work without a fresh prompt — triggered by an event, a schedule or their own judgement. After Google, Anthropic, xAI, Microsoft, Meta and OpenAI each shipped one between May and September 2026, the term hardened from a description into a recognised product category.^[1][2]

History

The idea is older than the products. Scheduled jobs and rule-based automation (cron, IFTTT) let software act without being told each time; the 2022 ReAct pattern gave models a loop of reasoning and tool use; AutoGPT and BabyAGI in 2023 chased goals autonomously; and 2025 brought task-completing agents such as Codex, Claude Code and Operator that work step by step but stop when the task ends.^[1][2] What was missing was the second axis — continuity. Technical users assembled it themselves: scheduled tasks, notes in files, separate agent accounts, and from late 2025 self-hosted personal-agent frameworks such as OpenClaw and Hermes Agent.^[2]

In 2026 the hosted wave arrived: Google shipped Gemini Spark at Google I/O in May, Microsoft introduced Scout in June, Anthropic launched Claude Tag as a persistent teammate in Slack on 23 June, xAI opened Grok Bot's beta on 11 August, Meta launched Muse on 8 September, and OpenAI released dots at DevDay on 29 September.^[1][2][3] Within a few months the shape converged: a persistent agent identity, its own cloud computer, connected apps, approval rules and a chat channel to talk to it from.

Key Concepts

• Continuity over capability. The agent manages its own memory — deciding what to keep and when to recall it — instead of starting from whatever the last prompt contained.^[2]
• Its own computer. Most hosted versions run in a sandboxed cloud environment that outlives the conversation, so work continues when the user closes the laptop.^[3]
• A control boundary. Products differentiate on what the agent may do alone versus what needs approval; sensitive actions such as sending money or changing a password are held back for the user, and credentials are kept in a secure store that the model itself cannot read.^[1][3]
• Waking conditions. An event in a connected app, a schedule, or proactive judgement — the three ways work starts without a prompt.^[1]

Applications and Impact

Documented uses include an agent that watches customer feedback for recurring requests and returns pull requests, one that prepares revised assets when a launch scope changes, and one that noticed an early tester had forgotten to invoice a publication and sent the invoice after approval.^[1] The category also carries an unusually loud critique: users report token bills running to hundreds of dollars from unsupervised sessions, agents that stall overnight, prompt-injection exposure when an agent has broad tool access, and credential concentration that makes switching vendors painful — pushing "guardrails and harnesses" toward being the enterprise requirement for the category.^[3]

>See Also

• Agentic AI
• Autonomous Agents
• AI Agents
• Multi-Agent Systems
• Prompt Injection
• Model Context Protocol