An agent being always online doesn’t mean it’s always running
A person can only follow 3–5 Agent Sessions at the same time. A design consensus around always-on is taking shape, but the real bar has not yet been crossed.
DeepSeek, in 'Super-agent Design Exploration,' notes that a person can usually follow only 3–5 active Sessions at once (independent working contexts). Agents can already do work on their own, but keeping that work moving still depends on human coordination.
While reviewing several recent designs, Li Jian AI found that DeepSeek’s Super-agent, Pi, LangChain’s Async Subagents, and Grok Bot are beginning to make some similar choices. He distills them into three design consensuses for Always-on Agents.
First, Always-on does not equal Always-running.
An agent that is online for the long term does not need to run continuously. Grok Bot users interact with a Bot that persists, but the Cloud Computer that actually executes commands can start, sleep, and resume on demand. Pi Durable achieves a similar effect with persistent Sessions. DeepSeek’s design separately manages the computation of the Agent Loop, the Session state that must be persisted, and the environment that actually executes code.
What persists is identity and state, not a process that runs forever. Recovery mechanisms are also layered: Memory stores experience, Durable Workflow stores progress, and virtual machine snapshots store the environment. Logs record tool calls, but that does not mean the external operations necessarily succeeded.
Second, work persists, while executors appear on demand.
The computing environment can exit, but tasks cannot be lost. DeepSeek’s Router / Worker architecture separates global coordination from concrete execution. The Router continuously receives messages, while Workers do the work in independent Sessions. More important is the Ticket—an authorized unit of work, with a Mandate and an Assignment attached. After a Worker exits, the Ticket remains, and the system can reassign the executor. If the user changes requirements midway, the Executor cannot expand its authorization on its own; it must return the matter to the coordination flow for a new judgment.
LangChain’s Async Subagents are similar: the main Agent can query status, add requirements, and cancel background tasks. More important than whether we call it a Worker or a Helper is whether, after an executor exits, the system knows how the work should continue.
Third, autonomous wake-up does not equal autonomous action, and still less does it equal notifying the user.
LangChain’s engineering Agent Robin watches Slack discussions but does not speak every time. The team introduced a decision model to judge when there is something worth saying. A sales Agent classifies email threads as 'pending,' 'draft prepared,' or 'no reply needed.' LangChain disclosed that more than 70% of runs come from scheduled triggers rather than user initiation.
An Agent seeing an event does not mean it must act; being able to act does not mean it should interrupt people.
The three consensuses point to the same question: how to make work no longer depend on a person watching a Session. State is independent of the process, tasks are independent of the executor, and the system coordinates proactively based on events. Ideally, a long-running piece of work can form a continuous loop: trigger → evaluate and authorize → execute → verify results → decide the next step.
But keeping work from breaking does not mean the system can already be continuously accountable. At least three questions remain unresolved: Can it continue? How does it recover after a crash? If an external operation succeeds but is not recorded, how do we avoid duplicate side effects? Should it continue? Is an authorization from several days ago still valid? Is it continuing in the right way? The work has not stopped, but it has drifted from the goal—who will notice?
DeepSeek proposes an independent Watchdog that not only checks whether a process is alive but also tries to determine whether a Worker has made substantive progress and whether it has drifted from the goal. A public prototype cannot yet show that these mechanisms have been implemented. As long as the system cannot judge whether work should continue and whether results are trustworthy, people will still have to keep checking and intervening.
Some developers have run into similar pitfalls in practice. geoqiao connected several project repositories to a manager Bot on Orca and, in less than 24 hours, generated 29 to-do items. He saw a real-time 'control room' panel that Danilo posted, with the top reading '33 items are waiting for your review.' People are not AI’s project managers. His summarized approach: split Workers by context and attach Skills by function; in daily operations, use one manager Bot with a WIP limit of 3–4 items; create a new Bot only when permission isolation, high-frequency silent checks, or memory conflicts require it. These practices align with the three consensuses—reducing the need for people to continuously watch Agents.
In the past, people organized an Agent’s work around a single model execution; now, they are beginning to organize each model execution around work that persists. From continuous execution to continuous accountability—that is the next bar for Always-on Agents.
发布时间: 2026-11-22 06:10