AI / Agentic Loops for GTM

A recipe for making an AI agent do one job over and over, reliably. Every loop has the same skeleton: a trigger (when it starts), one bounded task (what it does), a check (how it knows the result is good), and an exit (when it stops or hands off to a human). Think of it as a standard operating procedure that runs itself.

A chat answer is one-and-done and depends on you being there to prompt it, judge it, and re-prompt. A loop runs on its own trigger, checks its own output against a fixed standard, keeps only verified wins, and knows when to stop or escalate. The reliability comes from the structure, not from a smarter model.

Scheduled loops run on a clock (every night, every Monday) — good for monitoring and sweeps. Event/hook loops fire the instant something happens (a prospect views your proposal, a health score drops) — good for real-time responses. Goal loops keep working until an outcome is reached or a defined exit condition hits (contract signed, or deal marked dead) — good for anything that needs persistence.

It's the loop's category. The technical layers describe how a loop works under the hood: Reasoning (how the AI thinks through a problem), Harness (the scaffolding that runs an agent), Multi-agent (several agents working together), Learning (loops that improve from experience), Training (how models themselves are improved), and Memory (how agents remember). GTM and the lifecycle stages describe what a loop is for: winning and keeping customers.

The bowtie is a picture of the full customer journey. The left side narrows like a funnel — Awareness (they discover you), Education (they understand the problem you solve), Selection (they evaluate you against alternatives) — down to the knot in the middle, Mutual Commit (the deal is signed). Then it widens back out: Onboarding (they get to first value), Retention (they stay), Expansion (they grow). Cross-stage loops are plumbing that supports all seven.

The unit the loop repeats on. A loop that cycles over "one thumbnail" produces and scores one thumbnail per pass; one that cycles over "scheduled loop" runs once per tick of its schedule. It tells you the size of a single turn of the crank.

No. Well-designed loops have explicit authority limits: what the agent may do alone versus what needs a human sign-off. You'll see HOTL ("human on the loop") in some descriptions — a person supervises outcomes and gets pulled in when the agent's confidence drops or the action is risky (spending money, sending to a customer, signing anything).

27 loops carry their full instructions — click ▸ prompt on a row and paste the prompt into an AI agent tool (Claude, ChatGPT, or a coding agent). Rows without a prompt are techniques or designs: the description tells you the pattern, but you'd build it into your own stack (ours run on Trigger.dev with Supabase as the data layer).

Pick one small, annoying, recurring task — not your hardest problem. Find a matching loop with a ▸ prompt, run it manually a few times to see how the check-and-repeat rhythm works, and only automate the trigger once you trust the output. One loop that reliably works beats ten that half-work.

Every loop is mapped to where it works in go-to-market. Beyond the dedicated GTM layer, all seven bowtie-funnel stages — awareness, education, selection, mutual commit, onboarding, retention, and expansion — are GTM stages, so most of the 112 loops target a specific point in winning, closing, or growing a customer. The technical layers (reasoning, harness, memory, and the rest) are the plumbing those GTM loops run on.

Marketing automation follows a fixed if-this-then-that path a human wired up in advance. A loop uses an AI agent that decides what to do inside a bounded task, checks its own output against a standard, and repeats or escalates — so it handles judgment steps like qualifying, drafting, scoring, and researching that a rules-based workflow can't. Automation moves data; a loop makes decisions.

Not to start. 27 loops ship with a copy-paste prompt you can run today in Claude or ChatGPT with no setup. Turning a loop into an always-on system — its own trigger, memory, and guardrails — is where a builder comes in; ours run on Trigger.dev with Supabase as the data layer.

No — they replace the repetitive parts of the work, not the people. A loop runs a bounded task and hands off to a human when the action is risky or its confidence drops (HOTL — human on the loop). The point is to free your GTM team for the judgment and relationship work only they can do.