The cost of running cold email agent loops at scale just collapsed.
β
Iβm using the NEW Fable 5 model to take over cold emailβ¦
β
A bit of context:
β
Boris cherny, who leads claude code at anthropic, has been saying it for a while:
β
"i don't prompt claude anymore. i have loops running that prompt claude. my job is to write loops."
β
Every engineer i know is using this idea to ship code.
β
NO ONE is using it on cold email.
β
So I built it for outbound.
β
Because across our 50 clients, 30,000 inboxes, and 2M emails/mo that we sendβ¦
β
The actions that I see actually hurt campaign performance is NOT the sending.
β
It's the judgment that runs on TOP of the sending.
β
> diagnosing why a campaign's PCPL drifted above its healthy band
> rewriting copy before ESPs fingerprint the structure at volume
> rebuilding a list when the bounce rate climbs past 3%
> repairing infra when sending capacity drops below what the campaigns need
> mining replies and call transcripts weekly so the copy learns prospect language instead of guessing it
β
Most agencies still have a person doing all 5 of those, by hand, on a calendar.
β
That's the layer I automated.
β
So I wrote up the full blueprint for the self-optimizing cold email system we run internally.
β
What's inside this blueprint:
β
1. the PCPL benchmark bands per offer type so you know when something's broken (high-ticket service, mid-ticket, product/wholesale, enterprise)
2. the 6 components every loop needs, mapped 1:1 to the outbound stack (automations, worktrees, skills, connectors, sub-agents, memory)
3. the 6 loops we run on every account: daily health check, script improvement, reply learning, inbox/domain health, monthly angle mining, new campaign pipeline
4. the 3 memory files per client (LEARNINGS.md, STATE.md, Postgres) that turn 6 months of campaign history β selection pressure on every new draft
5. the 3 agent files where the WRITER never grades its own work (copywriter / standards-checker / deliverability-checker) and why fable 5's self-validation doesn't replace the maker-grader split
6. the 5 failure modes that kill these systems (copy thrash, token spend, garbage skills, comprehension debt) the 3-stage rollout we use even on new client accounts
7. defensive autonomy: the ONE loop allowed to act without approval (pause a burning domain instantly. waiting for a human costs the domain)
β
Want it?
β
β’ comment "LOOPS"
β’ follow me so i can DM the link
β
PS
β
Fable 5 is what makes long-horizon agent loops viable for high-volume cold email right now.
β
Yusuke kaji at anthropic put it this way: "at the highest effort, claude fable 5 reflects on and validates its own work. for us, that's what makes highly autonomous operations possible."