General agent concepts

Always on agents

An always-on agent is an AI agent that keeps working when nobody is in the conversation. It runs in the background on standing goals, schedules, or incoming events, and comes back with a result, a question, or a request for approval. OpenAI used the term for its dots agents on September 29, 2026. Google describes the same shape for its Gemini agent, announced in October 2026, which can be scheduled or set to respond to events and keeps running for hours or days. LangChain's Harrison Chase called the idea an ambient agent in January 2025.

It differs from a chatbot, which waits for a person to send a message, and from a scheduled job, which runs the same predefined steps each time. An always-on agent decides its steps as it goes, so two runs from the same trigger can do different things.

Because nobody watches each step, what it did has to be read from the record it keeps, and small errors have more time to compound. As of October 2026, products handle this with activity views, session logs, and approval rules for actions that change something.

12 questions

Answered, plainly.

Does an always-on agent's risk of overstepping grow with its task chain?Yes. OpenAI's own testing found doubling a dots agent's chained tasks from five to ten roughly doubled its flagged boundary problems, 8.6% to 19.7%.answer →How do you know what an always-on agent did overnight?An always-on agent's night shows in the record it kept: traces or session logs of each step, the changes it left in other systems, and the actions it paused on for approval.answer →How is an always-on agent different from a chatbot or a scheduled job?A chatbot acts only when you message it, and a scheduled job runs fixed steps on a timer. An always-on agent runs unattended and picks its own steps each run.answer →Should an always-on agent have its own identity and accounts?Once it holds a standing role rather than acting as you, yes: OpenAI's specialist dots and Google's coworker agents each get their own identity, so actions trace to the agent, not a person.answer →What can an always-on agent do without asking for approval?OpenAI's dots check risky actions live through auto-review; Google's Gemini agent grants access up front and reviews what happened afterward, with no per-action sign-off.answer →What is an always-on agent?An always-on agent works with no person in the chat: it runs in the background on standing goals, schedules, or events, and reports back or asks for approval.answer →Why do long-running agents fail?Small per-step error rates compound over a run with no end, and the agent keeps acting on its own record of events, so one bad assumption carries into every later action.answer →How do you stop an always-on agent from going off track?You tell it, mid-run: OpenAI's dots let you redirect work already underway without restarting it. Google's Gemini agent has no equivalent.answer →How do you test an always-on agent before letting it run unattended?Run it against real inputs with its actions blocked, log what it would have done, and compare that to what should have happened, before anything goes live.answer →How much does it cost to run an always-on agent?Nothing extra by default: OpenAI's docs say dot conversations don't count toward ChatGPT usage limits. Tasks it hands to Work or Codex use those products' normal limits.answer →Can a scheduled always-on agent miss an entire run with nothing to show it was skipped?Yes: an always-on agent's record only covers runs that happened, so a scheduled run the host never started leaves no log and no error, just a silent gap.answer →Can an agent violate a policy when every action it took was approved?Yes: a guard that checks one action at a time can approve a violation that only exists across several approved actions, since no single one looks wrong alone.answer →

Try Tessary on Cloud.

start for free or self-host

free · 10,000 traces a month · 1 GB · 30-day retention · $10 one-time credit for triage and RCA · no credit card