Skip to main content
This guide builds one real agent from scratch. It goes deeper than the quickstart: it explains what each stage produces and what to look at before moving on. The example agent posts a daily digest of failed payments to Slack.

​
Prerequisites

  • The jstm CLI, signed in. jstm whoami should show your workspace.
  • An interactive terminal for the planning chat.
1

Decide what the agent is for

Before opening the chat, be able to answer four questions. Joyvis will ask them anyway, and having answers makes the conversation short.
  • Who needs progress? The billing team.
  • What situation creates the need? Failed payments accumulate overnight and nobody sees them until a customer complains.
  • What outcome means it worked? The billing channel has an accurate digest by 09:00, every weekday.
  • What proves it? A message in #billing listing exactly the payments that failed since the previous digest.
That last answer matters most. It becomes the evidence the agent is judged against.
2

Plan it

Describe the agent, then let Joyvis interrogate the gaps:
Expect questions about which Slack workspace, what counts as a failure, whether to include retries, and what to do if Stripe is unreachable. Those answers become boundaries and error policies in the spec.Type / at the start of a message to search the skills catalog and reference one by name.When the plan looks right, press Ctrl+C. Copy the session id it prints.
3

Commit it

This creates the agent and prints its work tree. Read the tree — it’s the first chance to see whether the plan matches what you asked for.If the tree is wrong, resume the chat and keep refining. Committing again is cheap.
4

Build a package

Build compiles the committed spec into an immutable, versioned package. The receipt names the version.From here the version is fixed. Any change means a new version, which is what makes every later run traceable to exactly what ran.
5

Check it against the spec

Exit 0 means the package is clean. Exit 1 means at least one blocking problem, each one citing the spec clause it breaks.Most findings are mechanical — a missing field, a bad enum, a dangling file reference. Fix and re-run. For findings you’ve decided to accept, see Acknowledging flags.
6

Check it can actually run

Readiness answers a different question from conformance. Conformance asks whether the package is well-formed. Readiness asks whether the skills it needs can actually execute here — are the MCP servers reachable, are the credentials present.If there are blockers, the command exits non-zero and names them. Many have a targeted fix:
7

Dry-run it

The agent executes end to end with side-effecting actions held back. Nothing is posted to Slack; nothing is changed in Stripe.Progress streams story by story. Add --verbose to see an output preview for each step.
8

Read the evidence

run show gives you the run’s status, its stories, its output, and the exact versions that executed. ticket list gives the same work in business language.Now compare against the answer you wrote down in step one. Did it identify the right payments? Would the digest have been accurate? That comparison — not the absence of an error — is what tells you whether the agent works.

​
What you have

An agent that exists as a versioned package, is provably well-formed, can reach its dependencies, and has completed a full execution without touching anything real. It is still at Dry Run. That is where it stays until you decide the evidence justifies moving on.

​
Next

Promote to production

Move through Pilot to Live.

Connect a service

Give the agent real Stripe and Slack credentials.