Free Instant SEO Audit

Quick Answer

Always-on agents keep working across sessions by using five core primitives: an event inbox, timers, memory, permissions, and an outbox. Each run receives an event, completes a task, saves what changed, and feeds the next cycle, which lets an agent act without a person starting every task.

The architecture described by William Hu in his November 13, 2026 guide on the Claude Platform covers agents that run in the background, work across runs, and deliver outcomes to users.

How Does an Always-On Agent Run?

An always-on agent spends most of its time asleep and wakes periodically to do work. Each period of activity is called a run, and the agent receives an event, completes its task, and saves information about what happened.

Each run operates inside a session, which contains the transcript of messages and tool calls. When the run ends, the agent returns to sleep, and the cycle continues. A run can also set a timer that wakes the agent later, while a user reply returns as an event in the inbox.

Durable Task for AI Agents (2026)

Durable Task provides infrastructure for AI agents that run for hours, call external tools, and must survive infrastructure failures. The Durable Task runtime checkpoints every state transition, including LLM responses, tool call results, and control flow decisions, to durable storage.

When a failure occurs, execution resumes automatically on a healthy VM. Completed LLM calls are not repeated. The programming model supports resilient, stateful agentic workflows using standard programming constructs, including loops, conditionals, and error handling, in .NET, Python, Java, and JavaScript/TypeScript.

Durable Task for AI agents lets agents run for hours, call external tools, and survive infrastructure failures.

Which Primitives Make up an Always-On Agent?

The surrounding architecture has five components:

  • An event inbox that receives work from sources such as webhooks, scheduled tasks, and background jobs.
  • Timers that support self-set follow-ups and recurring schedules.
  • Memory that preserves instructions, notes, past messages, and task progress between runs.
  • Permissions that govern credentials, higher-risk actions, and tool calls that need user input.
  • An outbox that manages messages the agent sends to users.

Together, these primitives support a cycle in which an event starts a run, the agent loads memory, takes permitted actions, updates its records, and delivers the outcome.

How Does an Event Inbox Prevent Lost Work?

An event inbox is the source of truth for work waiting to be processed. Event sources send entries to the inbox, and a background worker reads pending events and starts runs.

The simplest inbox implementation has three components:

  • A durable list of events, usually a database table.
  • A handler, usually an HTTP endpoint, that writes events to the list.
  • A worker, usually a background service, that reads from the list.

Direct calls from event sources to an agent create two specific failure modes. Simultaneous events can start competing runs, and a crash can lose an event. The inbox addresses these by recording every event, tracking its status, and deduplicating entries through identifiers from the source.

Hu recommends acknowledging events quickly because some webhook providers treat slow responses as failures. The worker should pull the latest pending events when it wakes, and only one worker should be active at a time. Pending events can be batched into one run, such as when a user sends several messages in succession.

Events that continue earlier work, including a user reply, can resume an existing session. Other events start a new session. For a business website, consider the inbox wherever background work may arrive from webhooks, scheduled tasks, or background jobs.

How Do Timers Trigger Agent Work?

Timers come in two forms. A follow-up is a deadline the agent sets for itself, such as checking for a reply. A schedule is a recurring task on a known cadence, such as a daily newsletter.

Timers live in a durable store separate from the inbox, which holds events ready to be worked on. A follow-up and schedule can share a similar schema, but a schedule also tracks its recurrence.

A timer does not start a run by itself. A background scheduler polls the timer store at a fixed interval, finds items that are due, updates them, adds an event to the inbox, and wakes the worker. A completed follow-up is marked as done, while a schedule has its next occurrence recorded.

What Should an Always-On Agent Remember?

Memory gives an agent continuity across separate runs. Hu recommends keeping memory in its own store, such as a database, object store, or directory of files mapped to durable storage.

The guide identifies four types of memory:

  • User instructions: Instructions about agent behavior, loaded at the start of every run.
  • User notes: Learned details such as a user’s role, preferences, and corrections. A single file can hold early notes, while an index of several smaller files can be loaded first.
  • Past messages: A record of what the agent has already told the user, either in a separate file or through the worker loading the agent’s last few messages.
  • Task progress: A file for each ongoing task that records completed work and remaining work.

A rule of thumb in the guide is to store only what the agent cannot look up again. When information already exists in another system, such as an issue tracker, the memory entry can be a pointer to that record.

Before doing work, the agent loads its system prompt, tool definitions, and memory files. This initial context is carried into every run, so the guide recommends keeping it lean and using prompt caching to lower costs. If a session grows too long, older messages can be compacted or a new session can begin, and both approaches lose detail that needs to be stored in memory.

Dreaming is an optional addition that reviews past sessions to curate memories, replace outdated entries, and identify patterns across runs.

How Do Permissions Control Unattended Agent Actions?

Permissions decide whether an agent proceeds with a tool call, blocks it, or waits for approval. A run holds the user’s access and operates without direct supervision, which the guide calls the area where always-on agents carry the most risk.

Three permission components apply:

  • Credential store: A proxy outside the agent’s environment adds sensitive credentials, such as passwords or access tokens, to outgoing requests. The agent sends requests without the real credentials.
  • Auto mode: Classifiers review higher-risk actions and block those identified as dangerous. A person still needs to review the most important decisions.
  • Deferred tool calls: A tool call can pause for clarification or approval. The run ends, the session is saved mid-call, and the user’s answer arrives as a new inbox event that resumes the session.

While a deferred call waits, the worker can process other events. Claude Code supports auto mode and deferred tool calls, while MCP Apps can be used to present requests through an interactive card.

What Does the Outbox Deliver to Users?

Most agents stream responses as they are generated. Always-on agents usually work in the background, so the user gets the outcome and a few details, not each step. The outbox manages the messages that the agent sends to users.

A run can send a message to the outbox, decide that no message is needed, or use a deferred tool call to request permission or additional input. A simple send tool handles direct delivery. A formal outbox becomes useful when message pacing or recovery from errors matters.

The outbox is a separate table of outgoing messages, paired with a background dispatcher. The dispatcher reads pending messages, decides what to deliver, and sends them to the user. A deferred-call request can go through the outbox or appear directly in the interface.

Message volume also needs attention. Hu recommends building evaluations to find the balance between too many notifications and too few. The suggested starting point is for an agent to begin quietly and earn more attention over time.

What the Evidence Shows

Hu cites recent model improvements as a factor in the renewed interest in always-on agents. Newer models have shown better judgment in understanding intent and knowing when to check in, while intelligence costs have fallen. On one long-horizon coding benchmark, Opus 5.5 cost 78% less per solved task than Opus 5, compared with 40% less on typical workloads.

Research on wearable agents provides a separate set of findings. The VisionClaw study evaluated an always-on system on Meta Ray-Ban smart glasses with 12 laboratory participants. The system produced 13% to 37% faster task completion and 7% to 46% lower perceived difficulty compared with agent-only and always-on-only baselines. An autobiographical deployment study with four participants covered an average of 13.8 active days per user, 555 interactions, and 25.8 hours of use.

The Always-OnAgents survey reviewed a coded corpus of 435 works and found that research concentrates more heavily on accumulating and retrieving state than on governing, recovering, or relinquishing it. Its authors describe their Always-On Evaluation Protocol, or AOEP-v0, as a pilot contract for scoring state mutation and recovery obligations.

Key Considerations

  • Always-on agents are asleep between runs. They do not process tasks continuously without receiving events.
  • Permissions carry the most risk because an unattended run holds the user’s access and can call tools without direct supervision.
  • Memory should hold information that cannot be looked up elsewhere, with pointers used for records stored in another system.
  • Build evaluations to balance message volume. Too many notifications create noise, while too few can leave users without needed information. Start an agent quietly and earn more attention over time, without sending too many notifications.
  • The 435-work survey corpus was treated as a scoped map rather than an exhaustive census of always-on agent research.

What to Do Now

  1. List every event source that could start work on your website or business system.
  2. Record which events belong in the same run and which unique identifiers support deduplication.
  3. Map the user instructions, notes, messages, and task progress your agent needs between sessions.
  4. Write down the credential, auto-review, and deferred-call rules for each tool it can use.
  5. Define the outgoing messages users should receive and the conditions for sending none.

FAQ

How Does an Always-On AI Agent Work?

It wakes when an event arrives, completes a task inside a session, saves what changed, and returns to sleep. Events can come from user messages, webhooks, scheduled tasks, background jobs, or timers.

What Are the Five Core Primitives of an Always-On Agent?

They are an event inbox, timers, memory, permissions, and an outbox. The inbox holds incoming work, timers trigger future work, memory carries context across runs, permissions govern actions, and the outbox manages outgoing messages.

What Does an Event Inbox Do?

It acts as the source of truth for pending work. Event sources write to it, a worker reads pending entries, and each entry has a tracked status. The inbox also supports deduplication through source identifiers.

What Types of Memory Should an Always-On Agent Store?

The guide lists user instructions, user notes, past messages, and task progress. It recommends storing only information the agent cannot retrieve from another system and using pointers when data already exists elsewhere.

How Does an Always-On Agent Handle Permission Requests?

A deferred tool call pauses the run and saves the session mid-call. The request appears to the user, and the answer arrives as a new inbox event that starts a new run and resumes the saved session.

Sources

Share This Post

Do you want help with your SEO or website development? Then Contact Us Here.

NewSunSEO

Grow in Google and AI Search

★★★★★★★★★★5.0 from 17 Google reviews

AI-SEO, local search, paid ads and websites from one team, with offices on Long Island and in Myrtle Beach.

  • AI-SEO
  • Local AI-SEO
  • AI Consulting
  • PPC
  • Web Design
  • Web Development

Request a Free Proposal

Tell us about your business and where you want to grow.

We will reply by email to talk through your goals and what a proposal would include.

    We use your details only to reply to you and send what you asked for. See our Privacy Policy.

    BizScoreAI award: #1 Internet Marketing Service - Myrtle BeachBizScoreAI award: Top Rated Local SEO - Myrtle Beach