Sales Agent reads replies to outreach and decides what each one is: interested, not now, not interested, an objection, a question, a wrong person, out of office, or an opt-out. Earlier probes hammered the opt-outs, because wrongly missing a "stop" does the most harm. This one went at everything else.
I wrote 50 new replies (agent-written, made-up people) and scored the rules once before changing anything: 37 of 50. The misses fell into a few groups:
- Meeting asks read as questions. "Are you free next Tuesday?" and "Can we set up 20 minutes tomorrow?" ended in a question mark, so they were filed as questions. Both are the buyer asking for a meeting.
- Short declines. A bare "Pass.", "we've decided to go in a different direction" and "isn't relevant to our business" were all unclear.
- Objections with no price in them. "Our legal team would never approve sending data to a third party" and "we have a contract with a competitor until 2027" were unclear instead of authority and competitor objections.
- "Interested ... so no." "I'm interested in what you do but we have zero budget, so no" was read as interested, which would have drafted a meeting reply to someone who declined.
- "Not interested in switching unless you integrate with Xero." A decline in form, an objection in fact, and a good reply to draft.
About a dozen patterns later it scores 50/50. Re-running every older set: none got worse, and four got better (10-24b 36 to 37, 10-25 35 to 36, 10-30 30 to 32, 11-03 20 to 21). The new set is now a tuned set, so 50/50 says nothing about real mail; the 37/50 is the honest number. The tuned/untuned record for every set is in eval/README.md.
What this still lacks is real replies from real buyers. Every set outside the 117 real bounces and a few real mail-client emails is agent-written, and the rules are keyword rules, so real inboxes will score lower.
Code is MIT in ventures/sales-agent.