TLDR:
- start with business value delivery in mind, not code.
- give all internal data points to the agent.
- HITL is feedback.
- better have infra ready.
I have read LangChain's piece on how they developed an agent for their own sales team.
The following are my notes, noted from their notes.
1. Start with idea of success, not code.
Despite being LangChain, a company whose job is to create a framework for the world to write agents in all kinds of ways, they did NOT jump straight to code.
They first planned what value (aka - either a few % of revenue or sales team productivity) the agent will extract.
2. Connect the agent to all data record systems.
Give the agent as much as data as possible.
The team was successful in extracting value because the data points were already in place for LangChain.
So, the agent could be more intelligent and fruitful.
3. HITL
One last read by a sales rep before the agent finally sends a cold email to a lead.
This HITL check actually turned out to be a source of feedback for the agent.
Like, the human sales rep didn't approve the email or he did. This is crucial feedback data for the agent.
4. Infra is important.
Multi hour agentic loops, if not multi days, require proper infra.
Parallel subagents, reasoning, and agent state (aka memory) - harness engineering becomes important here.
Top comments (0)