LangChain says schedules drive more than 70% of its managed AI-agent runs
Its internal engineering account puts human attention at the center: agents need to remember feedback, choose when to speak and make approval requests easy to review.
Loading page…
Its internal engineering account puts human attention at the center: agents need to remember feedback, choose when to speak and make approval requests easy to review.
Listen to this story
LangChain’s November 17 engineering account describes how its internal agents handle work that runs without a fresh prompt. More than 70% of runs on Managed Deep Agents start from schedules, but that figure measures how runs begin—not whether they succeed. LangChain’s practical focus is keeping stored preferences current, making human review easy enough to generate feedback, and limiting interruptions to useful moments. The account suggests teams should judge proactive agents by approval time, downstream outcomes, and whether people act on their check-ins.
LangChain groups triggers into outside events, developer-set schedules, and schedules created by agents; Robin can start reviewing when a code-change request opens.
A scheduled “dreaming” job reviews recent activity records and updates memory by prioritizing useful information, retiring stale details, and merging duplicates.
LangChain built a dedicated interface for reviewing GTM Agent email drafts after finding that clearing a backlog in Slack was slow.
AI agents that start work without a fresh prompt still need something scarce: human attention. In a November 17 engineering account, LangChain said more than 70% of runs on Managed Deep Agents now begin with schedules. Its lesson from two internal agents is that useful autonomy depends on remembering feedback and knowing when not to interrupt.
Written by Sydney Runkle, the account draws on GTM Agent, which monitors sales leads and replies, sends daily briefs and drafts outbound emails, and Robin, which drafts and reviews proposed code changes. Both run on Managed Deep Agents. The scheduling figure describes how work begins; it is not a measure of how successfully that work gets done.
LangChain separates the triggers into three groups: outside events, such as a newly opened code-change request; schedules set by developers; and schedules agents create themselves. Robin starts a review when a new request appears. GTM Agent can create a daily briefing schedule after a user asks for one.
LangChain’s reported share includes scheduled runs rather than ad hoc human prompts.
When an agent wakes up on a schedule, there is no new prompt explaining what matters. It can fetch current information from an inbox or calendar, but LangChain argues that stored preferences supply a different kind of context: what this person wants, and what deserves their attention.
The company describes three inputs to that memory. Existing material can give an agent a starting point—for example, sent emails showing someone’s writing style. Edits and rejected drafts provide immediate feedback. Later outcomes, such as whether an email receives a reply, offer another signal about what works.
One approach the post describes is “dreaming”: a scheduled background job reviews records of the agent’s recent activity and updates its memory before the next run. The aim is not simply to retain more information, but to keep the information guiding future actions relevant.
The agents operate in Slack, where people can message them directly or tag them in channels. Scheduled reports return there too. But LangChain found that reviewing a backlog of GTM Agent’s outbound email drafts in Slack was slow, so it built a dedicated review interface.
That interface separates conversations waiting on a person from those needing a draft reply and those needing no reply. LangChain treats review speed as part of learning: easy-to-review work produces feedback, which helps the agent pick up preferences and judgment that may never have been written down.
Robin uses Jev, a decision model, to decide when to join a Slack conversation without being tagged. The intent is to contribute when useful rather than interrupt every exchange. Its updates start with a one- or two-sentence headline, with detail in a thread reply. Code-review requests include a short summary, the change’s size and a button to claim the review.
For teams assessing whether memory helps, Runkle proposes tracking outcomes: how long drafts take to approve, how often sent emails receive replies, and what share of agent check-ins people actually act on. That last measure puts the human cost directly into the evaluation: a check-in should earn the attention it asks for.
Loading discussion...
Join the conversation
Explain when an uninvited contribution would be helpful rather than intrusive.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.