Wink Pings

The Bottleneck for Proactive Agents Has Moved to the Human Inbox

LangChain reflects on proactive Agents: more than 70% of runs are triggered by schedules, memory determines judgment quality, and the review interface determines whether people get overwhelmed.

The LangChain team recently reflected on its internal experience building proactive agents. One number in the conclusions stands out: more than 70% of runs are triggered by schedules, while ad hoc human prompts account for only a small share. This shows that the way agents work has changed: they organize around ongoing goals and are driven forward by events and schedules, rather than only waiting for new instructions within a single conversation.

These systems have three types of triggers. Real-world events, such as new emails or new PRs. Developer-defined schedules, such as consolidating memory overnight or cleaning up stale PRs weekly. And schedules that the agent creates itself based on the task, such as following up the next day or generating a daily brief at the user's request. LangChain's internal GTM Agent is responsible for following up on sales leads, handling replies, and drafting emails; on the engineering side, Robin, formerly Open SWE, is responsible for drafting and reviewing PRs. Both are built on Managed Deep Agents.

When an agent is woken up, there is often no new human instruction. Connectors can provide the latest emails and project status, but user preferences, previous corrections, and working habits have to be pulled from memory. The original article divides memory sources into three categories: existing materials, such as emails sent in the past; immediate feedback, such as the reasons a draft was edited or rejected; and action outcomes, such as whether a sent email received a reply. The real difficulty is ongoing maintenance: what is worth remembering, what has become outdated, and how to handle duplicate or contradictory information. A common practice is called 'dreaming': scheduling background tasks to periodically review recent execution records, organize them, and update memory.

When measuring memory effectiveness, how much content is stored is of limited value as a reference. What matters more is whether drafts get approved faster, whether email reply rates improve, and how many of the questions the agent raises actually lead to action.

The human review experience is also treated by LangChain as part of the system. The GTM Agent categorizes conversations into waiting for human handling, needs a drafted reply, and no reply needed, and only sends the first category to humans. Robin uses Jev to decide whether it is necessary to join a Slack discussion, avoiding chiming in whenever it sees a message. Email drafts have a dedicated review interface, because handling pending emails one by one in Slack is very inefficient. Human edits are not just completing an approval; they are also feeding back to the agent preferences that are hard to write down as rules in advance.

LangChain 主动式 Agent 原文配图

There are several other design details. The daily brief is sent to Slack, and follow-up content returns to the same thread. Robin first gives a one- or two-sentence summary, then expands on details. When requesting a PR review, it includes a one-line summary, the size of the change, and a claim button, lowering the cost of taking it on. Together, these details determine whether a proactive agent is a reliable collaborator or just another tool that generates notifications.

In the comments, someone pointed out that tracking only the proportion of reminders that lead to action is not enough; you also need to see how many important items were missed. If you optimize only for adoption rate, the agent may learn to report only the safest things. The original four-quadrant diagram separates 'useful silence' from 'missed opportunities,' which is closer to genuinely being less disruptive than simply reducing the number of notifications.

A netizen also shared their experience running influencer collaborations with a bot: the hardest things to teach were which ones to follow up on and what tone the reply should take. These preferences are hard to write down as rules in advance and almost always surface when editing drafts. This experience explains why LangChain places so much importance on the review interface.

A more practical judgment is that proactivity in a proactive agent includes two capabilities: continuing to work without new prompts, and taking up less of people's attention while doing so. Events and schedules address the former; memory and the review experience address the latter.

For a good long-running agent, the number of runs is not the only metric. What matters more is whether each inquiry is more accurate, whether each review is easier, and whether feedback makes it into the next judgment. Greater autonomy should be built on this kind of continuous improvement. Simply reducing confirmation steps easily pushes the burden back onto people.

Original article link: https://www.langchain.com/blog/what-weve-learned-building-proactive-agents

Sydney Runkle's article: https://x.com/i/article/2122696441508270080

发布时间: 2026-11-19 00:30