Examples

Three production-grade agent templates you can fork, customize, and ship today.

Three real example agents live under /examples on the marketing site — running templates you can fork the moment your CEO asks for "an AI feature."

This page is the decision matrix: when to fork which.

The three examples

| Example | When to use it | Forks well as | | --- | --- | --- | | Stripe Triage Agent | You get many inbound events to classify and respond to | Customer support triage, intercom auto-reply drafts, internal alert routing | | GitHub PR Reviewer | You want first-pass code review that learns your team's style | Linter on steroids, "first reviewer" gate, hand-off helper for senior engineers | | Customer Research Agent | You want a structured dossier per company, cached across runs | Pre-sales research, competitor monitoring, market sizing, hiring research |

What they share

All three follow the same shape — and so will the next agent you bring online:

  1. Memory recall at the top — "have I seen this before?"
  2. Real work in the middle — browser, HTTP, shell, or a combination
  3. Memory write at the end — capture the lesson for next time

Once you've internalized that pattern you can write the fourth (and the fifth, and the tenth) from scratch in 30 minutes.

How to fork an example

Each example ships:

  • README.md — the pitch, the customer scenario, the ship-time math
  • agent.md — the full task prompt (~60-90 lines)
  • run.sh (production) and run-local.sh (localhost:3000)
  • expected-output.md — what to expect on a typical run
  • .env.example — the env vars the agent needs
bash
cp -r examples/stripe-triage examples/my-agent
cd examples/my-agent
# edit agent.md to your shape
# edit run.sh to rename the agent
chmod +x run.sh
./run.sh

The examples are MIT-licensed inside the repo (they're not part of the Jettson runtime — they're customer-facing reference material).

Why "real" examples and not hello-world

Anyone can ship a "what's the weather" demo. The hard part of building agent features is shaping the prompt, picking the right tools, and using memory well. The examples are deliberately three customer-facing scenarios — the requests engineering teams actually field — so the prompt patterns transfer to whatever you're building next.

Want a fourth?

If your scenario doesn't match any of the three, the closest one is still your best starting point. Email support@jettson.dev with what you're building — if it's general enough we'll add it to the gallery.

Related