Skip to main content
This tutorial builds one agent, order-support, in five parts:
  1. Try it: run it on this page. No account, install or API key.
  2. Build it: run the same code on your machine, then connect it to DecimalAI.
  3. Start from a template: give it the Support pack’s prompt and skills, and make one skill your own.
  4. See what it did: read its traces and versions.
  5. Measure it: run a set of cases against two setups and compare.

​
Run the agent

Open app.decimal.ai/try. The page shows the agent before it runs:
  • Its prompt: one line, You are Acme’s order support agent. Look the order up before you answer.
  • Three skills on its menu, from the DecimalAI registry: billing-dispute-resolution, account-deletion-verification-policy and diagnostic-triage-questions. A skill is a set of instructions the agent can load when a question calls for it.
  • One tool, lookup_order, which knows two sample orders: ORD-1001, an electric kettle, and ORD-1002, a mechanical keyboard that was charged twice.
Pick one of the three example questions, or type your own. Pick an SDK tab (the OpenAI Agents SDK is the default) and click Run.

​
What a run shows

The reply comes back with How it got there, the steps the agent took:
  1. Offered 3 skills: the menu the model saw.
  2. Called lookup_order("ORD-1002"), with what the tool returned.
  3. Loaded billing-dispute-resolution: the model asked for that skill’s full instructions.
  4. Replied, with the model, the time and the tokens.
Each example question needs a different skill, so asking another one shows the model choosing a different one. The page reports what the model did, not what we expected: if it loads no skill, the page says so.

​
What ran

The code panel shows the program that ran for the tab you picked, with your question in it: the set-up commands, then main.py. It is the code Part 2 has you run. On this page the agent runs keyless, without a DecimalAI key. Its prompt is the fallback= text in the code, its skills are the three listed in the project’s .decimal/skills.toml, and nothing is sent to DecimalAI: not your question, not the reply, no trace. The run is also capped at a few turns and tokens. The tabs are four ways to build the same agent:

​
Free runs

You get 3 free runs in this browser. A failed run gives its run back, and a reply that didn’t reach you is sent again rather than run twice. Decimal pays for these model calls, so they also have a daily limit per network and a total cap; when either is reached, the page says replies are paused. After your three runs, the same agent runs in your own code with your own provider key. That’s the next part.

Next: 2. Build it

Run the same agent on your machine, then add a DecimalAI key to trace it and edit its prompt from the dashboard.