AutoGTM

Karpathy's autoresearch loop, applied to outbound.

Clone it. Fill in the company. Hand AGENTS.md to any agent. It treats every campaign as an experiment: change one thing, send (you approve), wait, score what actually happened, keep the winner, discard the loser, write the lesson into a brain. Run it long enough and outbound stops being a static playbook and becomes a learning machine.

It is channel-agnostic and client-agnostic. Email, LinkedIn, DMs — same loop. Nothing in this repo is specific to a company or a vendor. You point it at a company in the config block of program.md. Company facts and tools flow through the loop as data.

The idea in one table

Karpathy's `autoresearch`	AutoGTM
Fixed eval = `val_bpb` in `prepare.py` (agent cannot touch it)	Fixed eval = `attributed_intent_quality` in `eval.md` (agent cannot touch it)
Modifiable artifact = `train.py`	Modifiable artifact = `campaigns/current.md`
Loop = train 5 min → keep/discard	Loop = send → wait the attribution window → keep/discard
First run = unchanged baseline	First run = unchanged baseline
Output = a better model	Output = a sharper GTM brain + an action queue

Outbound has three outcomes, not one

Most teams optimize reply rate on a single channel. That discards campaigns that are working.

Outcome	What happened	Role in AutoGTM
1. Reach	The message was deliverable. You still cannot prove they read it.	Guardrail. Opens are not a metric. Hygiene is.
2. Reply	They answered: yes, no, or unsubscribe.	Eval input (T0–T3).
3. Silent self-serve	They never replied. They Googled you, asked a model, looked you up, hit the site, and acted.	Eval input (S1–S3).

The one keep/discard number is attributed_intent_quality: weighted buying intent per reached attempt, with replies and silent conversions collapsed per person (no double-count). A low-reply campaign that books meetings from the site is a keep. Full ruler: eval.md.

The operator runs this as G.E.A.R. — Get → Evaluate → Author → Release — the RevOps name for the same autoresearch loop. Mapping lives in program.md.

Quickstart

git clone https://github.com/bcharleson/autogtm.git
cd autogtm

Fill the Configuration block at the top of program.md (company, ICP, offer, channels, CRM, analytics, the agent running it).
Open the repo in Grok Bot, Hermes, OpenClaw, Claude Code, Cursor, Codex, or any agent that loads AGENTS.md. If your host needs extra wiring, load one file from runtimes/.
Seed brain/ with anything you already know (optional).
Tell the agent to read AGENTS.md and start. You approve every send. It scores, logs, updates the brain, and asks for the next send until you stop it.

That is the whole product. There is no app to install.

Repo layout

AGENTS.md                   ← hand this to the agent. Auto-loaded by most runtimes.
program.md                  ← the one file a human edits (config + loop + rules)
eval.md                     ← the fixed ruler. The agent does not edit this.
methodology.md              ← why it is shaped this way
campaigns/
  current.md                ← the single artifact the agent edits per experiment
  results.tsv               ← one row per experiment (keep / discard / pending)
  ledger.tsv                ← one row per attempt (how channels collapse to a person)
brain/                      ← compounding memory. Ships empty. Typed entities, see brain/README.md
channels/                   ← email · linkedin · dm contracts (same loop, different tool)
runtimes/                   ← optional host adapters (Grok Bot, Hermes, OpenClaw, Claude Code)
loops/                      ← message, audience, objection, channel, silent-serve
deliverables/templates/     ← action queue + follow-up packet the loop emits

Lean on purpose, in the spirit of autoresearch. Three files matter: program.md (human), eval.md (fixed), campaigns/current.md (agent). Everything else is scaffolding.

What the agent may change — and what it may not

May change, one per experiment: audience, message, CTA, or channel.

May not change: eval.md, sample-size / attribution-window mid-experiment, sending infrastructure, more than one dimension at once.

A human approves every outbound send. The inner loop is autonomous. The send is not.

Runtime adapters

The loop is runtime-neutral. Adapters are optional glue (paths, chat formatting). Ignore them if you do not need one.

Runtime	Adapter
Grok Bot	`runtimes/grokbot.md`
Hermes	`runtimes/hermes.md`
OpenClaw	`runtimes/openclaw.md`
Claude Code / Cursor / Codex	`runtimes/claude-code.md`

Works with crystallized-intelligence

AutoGTM's brain/ is crystallized-intelligence-compatible. AutoGTM generates and scores the experiments that fill the brain; crystallized-intelligence compiles that feedstock into tight agent-readable layers. You do not have to use both — a plain folder of markdown is enough.

A named-persona wrapper that embeds this loop lives at gtm-operator if you want an identity file on top. You only need this repo to run AutoGTM.

License

MIT — see LICENSE. Clone it, fork it, run it on any channel, sell what you build with it. A gift to anyone whose outbound should learn.

bcharleson/autogtm

README