LangChain and Actively AI moved sales agents from one per account to one per job, because reps couldn't review them efficiently
How you scope an agent decides how easily people can correct it, and Actively AI rebuilt its product around that.
What happened
LangChain CEO Harrison Chase, with Mihir Garimella and Anshul Gupta of Actively AI, published a post on what they call goal engineering. Their argument is that agents fail mostly from missing context, and the hardest context to supply is the kind nobody has written down. That knowledge has to come from human experts as feedback, so an agent's scope should make feedback as easy as possible to give.
Actively builds agents for sales teams. Its first version ran one agent per account, with the goal of moving that account forward and producing actions for reps to approve. Two problems showed up. Reps switched constantly between agents that each had only a few items to review, and lessons about outbounding learned on one account had to be copied to every other account by hand.
The second version, built on LangChain's Deep Agents, is organized around "motions". A motion is a job such as outbounding or early-stage deal progression, running across many accounts. Reps review similar tasks together, and feedback on one task applies to the rest of that motion. Separate background agents gather each account's context, and the motion pulls it in when needed. The post doesn't share performance numbers for either version.
My take
This fits how CRM automation tends to work. Account and contact data is usually already in the CRM and fairly objective. What the system lacks is the team's judgment: how aggressive a follow-up should be, when to escalate, which phrasing a manager rejects.
If I were scoping a sales agent today, I'd split it the same way. A context job keeps each record's summary current. Separate job agents handle follow-up drafts, renewal reminders or meeting prep across the whole pipeline, and each writes into a review queue grouped by job type. Approved and rejected drafts become shared instructions for that job.
That design also makes evals easier, because every task in a queue shares the same success criteria. A reviewer's time is the real constraint here, so I'd budget for the review screen early.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
