Day One sessions

The last session extends the exact repo from session three. You pull a prepared branch, so nobody spends the evening on plumbing. We build a support agent that answers from your own content and knows when to hand off to a human, write the guardrails before it goes live, then put one workflow on a schedule and watch a run land in the logs. Then we break it on purpose, because a job that failed silently is the worst outcome available.

Dates announced soon — evenings, 6:00 PM at Workuity Biltmore. Watch Discord and the Luma and Meetup calendars.

Format: the host presents and builds live on screen. Bring a laptop and work along if you want to, or just watch — nobody is put on the spot or asked to present.

Why this session exists

An agent that's confident and wrong costs more than no agent at all. It reaches for model memory instead of your pages, promises things you would never promise, and does it in front of a customer. The fix isn't a better model. It's a guardrails file, three legal outcomes, and a human at the end of the queue.

What we cover

1. 01

   Agent vs workflow

   One waits for a message, one waits for a clock. Different triggers, different shapes, different ways of going wrong. We build one of each and name which of your problems is which.

2. 02

   The handoff

   Pull the prepared branch onto session three's repo and add your keys. Ten minutes of wiring somebody else already did, so the evening goes on the parts that need your judgement.

3. 03

   The support agent

   A chat route on your own site, answering from your own pages rather than model memory. Three legal outcomes: answer, escalate, refuse. "I don't know — here's a human" is a correct answer, and making it correct is most of the work.

4. 04

   Guardrails, written first

   Never promise a date, never negotiate a price, never send email. Written to a file before deploy, because the version you write afterwards is the version written after the screenshot.

5. 05

   The night shift

   One scheduled workflow, picked in the room: weekly article drafts into a review queue, or the numbers report rebuilt to run itself. Nothing publishes without a human. The path we don't build is written up in the guide.

6. 06

   Watch it run, then break it

   Manual trigger first and read the whole log. Then enable the schedule, wire a failure alert, and make one run fail on purpose so you know what the silence would have hidden.

7. 07

   What compounds from here

   Four sessions ago this was a blank box. Now something runs while you sleep. Where it goes next is a business with numbers on it, and that's the job Zero to Launch — the six-session program and the hackathon weekend — picks up.

What to do afterward

Nothing is required. The program compounds if you run the same work on your own product; this is the suggested output for the session:

04-agents/ — if you want somewhere to put it: your guardrails file, the agent's ten-question test log including the two escalations, the workflow, and a screenshot of the run you made fail.

Everything is open source. Missed the session? It's archived in Discord.

The session kit

Slides, a follow-along guide and skill files, free and downloadable whether you make it to the room or not.

Install this session's skills

curl -fsSLO https://www.shipai.club/skills/day-one-04-agentic-workflows.zip
unzip -o day-one-04-agentic-workflows.zip -d .

What's in the bundle

• /support-agentSKILL.md

  Put a chat agent on the site that answers from your own content and knows when to hand off to a human. Use on the prepared template branch.

• /scheduled-workflowSKILL.md

  Ship one workflow that runs on a schedule and lands its output in a review queue. Use once a routine is worth taking out of the week.

Every skill from the program is on the skills page.