Scheduled work now accounts for more than 70% of runs on its Managed Deep Agents platform, shifting the design problem from answering prompts to deciding when an agent should act or interrupt a person. Its internal sales agent and coding agent both wake for events and schedules, then report through Slack.

That autonomy makes memory operational rather than optional. LangChain separates useful signals into prior context, direct feedback and later outcomes. It warns that simply saving everything creates stale or contradictory records, so background maintenance must rank important facts, expire old ones and merge duplicates.

The company also treats human attention as a limited resource. Its sales system sorts conversations by whether they need review, while the coding agent uses a separate decision model to judge when it should enter a Slack discussion. Requests for approval are kept short and linked to focused review interfaces so corrections can flow back into memory.

These are lessons from LangChain’s own deployments, not a controlled comparison of agent architectures. They point to concrete evaluation measures, however: how often people act on an alert, how much they edit a draft and whether later outcomes improve. An agent that wakes frequently but produces ignored notifications is not becoming more autonomous; it is shifting its workload onto the reviewer.