Skip to content

Deliver with agents

In short

  • With a builder token, your own agent delivers a model through ParleML: frame, data, baseline, iterate, challenge, decide, release.
  • The agent writes no ML code. Features are recipe SQL, experiments are reviewed plans, and every number comes from ParleML.
  • You answer each decision. Promoting the model is your click in the app.

With a builder token, the agent you already use — Claude, Codex, Gemini, Cursor — delivers a model through ParleML's own engine. It frames the problem with you, shapes the data, runs and compares experiments, challenges the best one and registers it for your release. The agent writes no ML code: features are recipe SQL, experiments are reviewed plans, and every number comes from ParleML. You make the decisions.

Set it up

  1. Mint a builder token in Settings → Access: pick the builder scope, the workspace(s) it may build in, and its run budget (concurrent runs, runs per day). See Tokens & scopes.
  2. Connect your agent with that token (Connect your agent) and install the skills (Skills): ml-delivery is the one for this.

Run it

Use the deliver_model prompt (or ask in your own words):

Deliver a churn model in the Retail workspace: who will stop ordering in the next quarter?

The agent starts from get_workspace_brief: the workspace's datasets, untriaged leakage, the models in production and shadow, running and recent experiments, and next-step hints. A typical delivery:

Step What the agent does What you do
Frame Proposes the target, the population, the prediction moment, the metric and the costs, asks your requirements (batch or real-time, explanations, compute) and proposes one template for them — its recommendation first Confirm or change them; answer its open questions
Data Writes the training table as recipe SQL over the knowledge base, with its prediction moment; ParleML checks every column against it. Sets column roles; triages leakage findings with a reason for each Approve the table; confirm the leakage triage; answer its open questions
Baseline Plans an experiment on the approved template, states its hypothesis, launches the reviewed plan Say whether a doubtful top driver is known at prediction time
Iterate Changes one data lever per experiment within the template, compares with the parent and the model in production on validation evidence, and mines a promising run's rules, archetypes and text terms for the next idea —
Challenge Reviews the best candidate as a sceptic and ends with a verdict: pass, pass with conditions, or block Read the verdict
Decide Simulates operating points with your costs, and each condition's fallback Pick the operating point; confirm each condition, or take its fallback
Release Registers the model with its AutoDoc Promote it in the UI, when you are ready

Each plan records who made it — the token, the MCP tool and the host.

What the agent cannot do

  • Promote a model, archive or restore a workspace, delete anything, or touch settings — for every token scope. An archived workspace is read-only: every build in it is refused (409 workspace_archived) until a person restores it. A release happens only when a person promotes the model in the UI.
  • Train on the future: ParleML checks recipe SQL against the prediction moment the recipe states (or each row's as-of column) and refuses SQL whose columns read rows dated after it — an aggregate with no time bound, a join that can pull later rows, a window that looks ahead. You see the same findings as warnings on the recipe in the knowledge base, where a person's SQL is never blocked.
  • Retrain in place: training again is a new, reviewed plan, within the token's budget.
  • Build outside its workspaces, or past its run budget.
  • Run SQL on your source database, or read files on the server: recipes are its own SELECT on the sandboxed DuckDB engine.

Answering the agent

The agent asks in its own chat, with its recommended option first; your answer, in your own words if none of the options fits, is its next instruction.

Next