What are long-running agents? Explained with a night shift's handover notes

On 5 November, AWS made Kiro bots generally available: named AI agents that keep working on a goal over days, with their own memory and schedules. Here is what a "long-running agent" actually is.

In one sentence: A long-running agent is an AI agent that works on one goal over hours or days, across many separate work sessions, by saving its progress outside the AI model and picking up from those notes each time, instead of finishing everything in a single conversation.

The everyday picture

Think of a hospital ward, a hotel front desk or a factory that runs day and night. No single worker stays for the whole week. Work is done in shifts.

The night-shift worker arrives knowing nothing about the day. So the day shift leaves handover notes: what is finished, what is half-done, what to check first. At the end of the night, the worker updates the notes for the morning. A manager reads the notes and signs off anything important.

The worker is the AI agent. One shift is one work session. The handover notes are the progress the agent saves. The manager is you.

How it works, step by step

1. You set a goal, not a single task (the job board for the whole week). For example, "keep our build healthy" or "work through the bug list". Kiro's documentation describes a bot as an agent that takes "ongoing ownership of a goal".
2. The agent works in sessions (shifts). An AI model can only look at a limited amount of text at once, so a big job is split up. Anthropic's engineers describe the core problem plainly: each new session starts with no memory of the last one.
3. Before stopping, the agent writes its progress down (the handover notes). In Anthropic's approach, the agent keeps a progress file, a list of every feature with a pass or fail mark, and saved versions of its work with clear descriptions.
4. The next session reads the notes first (the new shift checks the board). It looks at what changed, checks that the basics still work, then picks one next piece of work.
5. The agent reports back and asks before risky steps (the manager signs off). Kiro's example deployment bot asks for your approval before each change, and a bot cannot gain new tools until you approve the request.

Some agents also run on a timer, like a scheduled shift. Kiro's example bot checks the build every 30 minutes and reports what it finds in one chat.

Where the picture stops

• A night-shift worker still remembers their own training and last week. The agent remembers nothing between sessions except what was written down and read back.
• Real shift workers notice when notes look wrong. Anthropic found agents can trust the notes too much, for example seeing lots of progress and declaring the job done too early.
• One worker covers one shift. A long-running agent can start several helper sessions at once and keep track of all of them, which Kiro bots are built to do.

Why it matters for your business

• Speed: slow, repetitive jobs, such as checking failed builds overnight or clearing a backlog, can keep moving while your team is away.
• Less babysitting: instead of restarting a chat and re-explaining the job, the agent picks up from its own notes.
• Risk: an agent that runs for days can make many changes. Approval steps, limited permissions and a readable log are what keep that safe.
• Cost and limits: long jobs use more computing. Platforms set limits; AWS's AgentCore Runtime, for example, runs sessions for up to 8 hours on its standard option, or up to 14 days on dedicated instances.

An example

Imagine a 15-person online shop whose website code is tested automatically every night. Tests often fail at 2am, and nobody looks until 10am. They set up a long-running agent with one goal: keep the nightly tests green. Each night it reads yesterday's notes, checks the failures, writes up the likely cause and drafts a fix. In the morning a developer reviews the drafts and approves the good ones. The agent records which fixes were accepted, so the next night starts from there.

Frequently asked questions

Is a long-running agent the same as an AI agent?

It is a kind of AI agent. Most agents finish one task in one conversation. A long-running agent keeps working on a goal across many sessions, using saved notes to carry on.

Does a long-running agent run nonstop?

Not usually. It works in sessions that start on a schedule, when something happens, or when you ask. Between sessions, its progress waits in storage.

How does it remember what it did yesterday?

Through notes kept outside the model, such as a progress file, a task list, saved versions of its work or a memory store. Kiro lets you view, edit or delete what a bot remembers.

Is it safe to let an agent work unattended?

It can be, with limits. Give it only the tools it needs, require approval for changes that matter, and check its work log. Its notes can be wrong, so people should review results.

Related explainers

• What is an AI agent?
• What is AI memory? Explained with an assistant's notebook
• What is a context window?
• What is human-in-the-loop?

The DevDrift builds AI automation with approvals and audit trails: see how.

Sources

• Kiro: Introducing bots in Kiro Crew (5 November 2026)
• Kiro Docs: Bots
• Anthropic: Effective harnesses for long-running agents
• AWS Docs: Host agent or tools with Amazon Bedrock AgentCore Runtime