LangChain's CEO says always-on agents bring new unsolved problems: when to notify, what to remember and how to stop goal cheating
Agents are moving from answering requests to running in the background. LangChain's CEO lists what nobody has solved yet.
What happened
In a new post, LangChain CEO Harrison Chase says agent builders have converged on standard context engineering tricks: summarizing long threads, offloading large tool results, skills for progressive disclosure, tool search and subagents. But every builder he has talked to lately is working on proactive, always-on, goal-oriented agents, and those bring new problems without general solutions.
His list: being proactive without being annoying, since too many notifications overwhelm users. Memory, where the hard part is deciding what to remember. Async work, including how to notify the main agent when background work finishes or it is paused waiting for human approval. Goal tracking over long horizons. Alignment, because agents optimizing a goal may cheat, as in deleting a codebase to remove all bugs. And browser use, which is harder to drive than a terminal.
The same day, LangChain updated its Managed Deep Agents post to say the product is in public beta on LangSmith Cloud. Memory, cron schedules, Slack channels, sandboxes and identity are files you add to an agent project, and one command deploys it. LangChain says its own Content Studio code shrank by about 80% after moving over.
My take
For client work, the first two problems matter most. A background agent that pings a sales team ten times a day gets muted within a week. Before building one, I would decide:
- What event is worth a notification, written as a rule, not left to the model.
- What the agent may remember across runs, and what it must forget, especially personal data.
- What a goal means in numbers, plus what the agent is never allowed to touch to reach it.
- Where it waits for approval, and who gets asked.
Tooling now makes deploying a scheduled agent easy. Deciding when it should stay quiet is still your job.
More posts
- Uber says AI agents booked four rides in its busiest week of October. Its app takes about 50 seconds to book oneNov 5, 2026
- A 2B parameter model beat frontier LLMs at pulling fields from invoices and IDs in its makers' benchmark, at 6 to 147 times lower costNov 5, 2026
- The best AI agent finished 29 of 45 real AWS tasks. About 80% of failures came from stopping too early, not from missing knowledgeNov 5, 2026