Learning Loops

This is what sets GTM Goat apart from a plain spreadsheet or toolchain setup: every interaction in GTM Goat pays into your Knowledge. A campaign that’s running isn’t just executing — it’s learning. Two loops run side by side, and both land in Knowledge.

Loop 1 — You and the craft

So the system doesn’t repeat the same mistakes and gets better over time at the build itself. This loop learns how work gets done in your workspace: your preferences and corrections, how to set up end-to-end campaigns with tables and references, which tool for what, how to build and debug an agent chain. It lives in Memory (Speicher) — self-maintaining markdown documents that agents read into their context and write forward after a session; a hygiene routine merges duplicates and fades what’s stale.

Loop 2 — Revenue

So the winning pattern gets replicated. This loop analyzes the GTM plays and derives structured Learnings: which offer, which ICP, which signal, which persona, which message, which channel actually books meetings and generates replies — not just delivers. What the system already collects for this today:

• Replies — the reply agent classifies every reply; thread insights extract which pattern led to a meeting or a decline.
• Positive signals — a reply recognized as positive gets analyzed and fed back into the qualification and copy steps.
• Copy and playbook adjustments — the closed-loop agent (Observer/Optimizer/Measurer) derives concrete proposals from reply patterns: switch copy, adjust cadence, sharpen a playbook — every proposal carries a confidence score, measured 48 hours later against the actual outcome.
• ICP and persona performance — which audience definition, which offer, which message leads to replies and meetings.

How a Learning arises and takes effect

The system distills generalizable, typed Learnings from the evidence of success (booked meetings, strong engagement, rated and corrected replies) — not a block of free text, but structured records with pattern, outcome, confidence, why it worked, evidence count, and source conversations. Every distilled Learning starts as a proposal and lands in an approval queue inside Knowledge. A Learning carries a confidence score that grows with further evidence and fades over time without new confirmation (reinforcement/decay).

Before every scoring or copy step, the table loads the approved Learnings into the prompt — so the same campaign learns while it’s running, not only in the next version. Because Learnings are structured, prompts, agent chains, and analyses read the same records instead of re-interpreting a block of text.

Always through approval, never silently

A Learning never silently changes a playbook or an asset. It proposes a revision — a copy switch, a threshold adjustment, an asset version — and that revision goes through the same Approval as any other structure-changing action. You see what the system learned in the queue, and approve or discard it before it takes effect.

The effect compounds over time and across campaigns: every new campaign launch in your workspace starts with the Learnings of the previous ones, not from zero.

Rolling out

The structured Learning assets with an approval queue are built as a first stage and togglable per workspace; distillation from evidence of success and self-maintaining Memory roll out next. Cross-workspace patterns (“what works across all workspaces”) stay deliberately separate and never flow into a single workspace’s prompt without asking.