The public builder is new. A hosted trial is available only when the deployment has configured its model access and trial budget; the builder explains when it is unavailable. This tutorial does not claim an external timed pilot has completed.
See the recorded example
The pack page shows an eligible recorded comparison when one is available. Read the task, both answers, model, date, and tested skill version. The example measures one skill on a selected case; follow the full benchmark link to inspect other outcomes. It does not establish quality for the whole pack. Missing or expired evidence is shown as unavailable, never replaced with a simulated live answer.Build and run without a DecimalAI account
Choose Start with 8 skills on the pack page to open the public builder. Review the prompt and selected public skills, edit the agent name, then choose Build locally. Creating the ZIP makes no model calls and creates no workspace agent. To get the same project from a terminal instead, use the CLI (decimalai 0.13.8 or later). It writes the project folder directly:
gpt-4o-mini-2024-07-18 as its default; it does not include a local-model adapter. Downloading it from the builder needs no SDK at all.
The ZIP includes agent.py, snapshot.json, exact skill bodies and text companions, requirements.lock, .env.example, bundle-manifest.json, and a README. The snapshot freezes all included content. Required missing companions or changed versions stop export. The runtime verifies file hashes before calling the provider. .env.example is a reference; export its variables in your terminal rather than assuming it is loaded automatically. Keep the key private.
--check-setup verifies files without calling a model. --checks makes four BYOK calls: two sample tasks and two semantic judge calls. It checks a useful support draft and an unavailable-action boundary against the actual delivered instructions. Receipts live in receipts/ on your machine. Missing credentials, incomplete provider output, and invalid judge responses leave checks incomplete or failed. A judge pass is a fallible assessment of these two cases, not production certification. Local receipts contain task and answer text; they are not uploaded.
Review the replies even when calls complete. In local verification with archived public Support skills, the default model still offered unavailable account actions, and the boundary check failed. The runtime has no tools to perform those actions. Stronger instructions alone did not qualify this model and skill combination.
To change the frozen configuration, regenerate the project from the builder or the CLI. A local successful run is a local completion. To inspect workspace traces, use the connected project path below.
Save the draft and try it live
From the same builder, choose Sign in to save and try live (Save and try live when you are already signed in). Sign in, return to your preserved draft, and explicitly save it into the displayed workspace. Canceling sign-in returns to the draft. The server revalidates the public selection and pins the reviewed prompt and skill versions; repeating the same save does not create another agent. If files changed since your download, review the changed starter before saving it. When enabled, choose one of the fixed sample scenarios and Run live sample. The initial trial uses the frozen Support instructions and a fixed OpenAI model on DecimalAI’s servers. It has no action tools or model judge. Its answer is ungraded; read it and inspect the linked trace. The UI shows the delivered versions/hashes and remaining allowance. The trial permits three lifetime attempts per signed-in identity and per organization. Failed or interrupted attempts consume allowance because a provider may have billed them. A global budget also bounds total sponsored use. Provider failure remains a failure; a recorded example never appears as a successful live run. Continue with your own model access after the allowance is used. A hosted trial does not verify your computer, local project, tool integrations, or deployment. For a connected runtime and exact local-run trace verification, continue below.Connected project: useful replies and matching workspace traces
This path creates a workspace agent, generates the SDK-integrated project, runs support tasks locally, and opens the exact traces produced by those runs.1. Review the Support pack and create your agent
Open the Support pack. Review its current starter selection. The packaged Support checks require these three skills; the pack may also select additional starters:
On the Support pack page, choose Create a connected agent (it opens the connected Support creation form), sign in if needed, and name the agent
my-support-agent. Review the selected skills and create it. If you already saved this agent through the public builder, reuse it instead of creating a duplicate. Keep all three check-required skills selected. If one is unavailable, resolve that missing selection before continuing.
Creation saves a prompt and skill assignments. Run the generated project in the next step to produce an answer.
2. Generate the project
Use Python 3.10 or later and a current published SDK:.env.example. Keep that file private. The default generated Support project uses LangChain; use the model shown in your setup instructions or pass a supported --model override when generating it.
The project contains agent.py, check_agent.py, checks.json, pinned SDK requirements, and a README. Run the actual generated agent:
3. Check the answer and open its trace
The check runs the same agent twice: a billing authorization-hold example and an account-deletion boundary example. It checks the actual skill bodies delivered to the model, resolved configuration, answer behavior, and receipt of the exact trace IDs. A model judge evaluates the supplied rubric; read its reasoning and the answer. Exit code zero means these two cases passed. It does not establish general support quality or prove that a skill improves every model. The included judge calibration applies to the documented OpenAI default; a different model/judge combination needs its own qualification. Open the dashboard link from the result and then View billing reply trace. Match its ID tocheck-results.json. Each attempt also retains check-results.<check-id>.json, including failures.
If the answers passed but trace confirmation is delayed, retry confirmation without new model calls:
4. Return with new customer work
Runagent.py again with a different ticket, amount, missing fact, or escalation condition. Reuse the project and agent instead of creating an agent for every ticket. Review each draft before sending it outside your application.
Change the prompt or skill assignment in the dashboard when your policy changes. The prompt is read when the process starts; skill bodies resolve at runtime. Restart the process and rerun the checks after changes. Pin prompt and skill versions when you need repeatable configuration, and requalify changes to tools, models, dependencies, or code. Hosting and tool integrations belong to your application.