Argus & Iris

The problem

Outbound involves a surprisingly large amount of work before anyone actually speaks to a prospect: identifying accounts, researching whether they're relevant, finding the right people, checking whether they've been contacted before, deciding which campaign makes sense, and getting them into it. At Nettle, most of that work was manual.

The approach

I built two agents to take that work over, and I talk to both of them in Slack. Argus finds the people worth contacting. Iris takes the ones who pass review and gets them into the right outreach campaign. I kept them separate on purpose, because deciding who deserves a message is a different job from sending one. They also never message each other. They share one record of every prospect, and the decision on a person is the handoff: Argus's work ends there and Iris's starts.

How it works

I give Argus a target, like a company, a person, or a conference attendee list, and tell him what I'm looking for. Before doing any research, he checks what we already know, so we don't research or contact the same person twice by accident. Then he researches the company, works out whether it's a fit, and looks for the people who could actually push a deal forward. Everything he tells me about a person comes with a source and a date, or he says it's his own judgment.

Once he has a shortlist, he sends each proposal to a separate reviewer agent that hasn't seen any of his research. Its job is to read the proposal like a skeptical colleague and ask one question: does the reasoning say something specific about this person, or would it be true of anyone with the same job title? If it's the latter, the person is dropped, with the reason.

When someone is approved, Iris checks that their record is complete and whether we've contacted them before or already know them through a deal. She then proposes which La Growth Machine audience they should go into (that's our outbound tool) and explains why. This is the one point where I step in. I can confirm the audience, correct it, or tell her not to add the person at all, all in plain English. And if I'm curious about someone the reviewer rejected, I can ask Argus to take another look and overrule it.

Key decisions

• Separate research from outreach, with a reviewer that has no memory of the research checking every proposal in between.
• Automate 80% of the system while creating specific gates to keep consequential decisions human-approved.
• Make every claim traceable to a source and a date.
• Design background work so it can fail visibly.

Current use

Argus and Iris are in use at Nettle and they work as intended. Argus turns a target into a short list of people with the evidence behind each one, the reviewer filters out the weak proposals, and Iris gets whoever passes into the right outbound audience. The only thing I do is check that Iris picked the right audience.

What I’ve learned

I learned that reliable autonomy takes much more than a capable model. Getting an agent to do the work is one thing. Making it do the work reliably, and knowing when it hasn't, took me far longer than I expected, including a full rebuild. It meant tying every claim to its source, keeping a human on the decisions that matter, and making failures obvious instead of silent. I also stopped trying to script what the agents say. I let them talk freely and put the rules in the data underneath.