Build and maintain investment models with an AI agent

Use this workflow when you want an agent to help research assets and maintain an investment model whose code, inputs, runs and allocations remain accessible after the conversation. An allocation is how investments are divided. ENSEMBLE supplies the persistent platform; the agent helps operate it.

Connect and establish the boundaries

You need an ENSEMBLE account and a compatible remote MCP client with connector access. Follow the current client setup guide, then authenticate through OAuth or the supported bearer-key configuration. Keep keys and private model code out of shared transcripts.

Ask the agent to discover the current tools, verify your account with get_account, check get_billing, and read get_runtime for models. Builds and runs are metered, while reads are free. Start with $100 in credit, valid for 7 days, with no card required. Builds and runs are metered. Check pricing for current terms. A spending limit is a constraint, not permission to keep retrying a failed charge.

Research and create one model

The sequence below is illustrative, not a reconstructed customer conversation. It follows the same asset selection and model inspection steps as the DIY walkthrough.

Inspect SPY and TLT, including coverage and missing data. Create a long-only model with 60% equities by default, the remainder in TLT, and an exposed monthly/quarterly/annual rebalance input. Make its execution assumptions explicit. Keep it private and leave automatic updates off until I review it.

Use list_assets, get_asset and get_asset_history to inspect candidates. The tools can return price history; coverage depends on the asset. Use create_model for the requested model, then get_model to inspect progress until its run is null. Check success or errors rather than treating an idle model as proof of a successful build.

The public 60/40 reference exposes equity_weight and rebal_frequency. Those names were verified October 5, 2026; read your own model's inputs_schema instead of assuming every build uses them.

Inspect before comparing

Use get_model for inputs, readiness, dependencies and the live run identity. Use get_build with its build_id to inspect your model's code; source access is owner-only. Read the run with get_run, and request the relevant blocks from get_model_history, such as metrics, allocation and latest.

Check dates, price basis, backfills, execution timing and costs before summarizing results. Separate observed history from synthetic extensions. Saved records support investigation, but do not guarantee immutable snapshots of every underlying data revision. See the methodology.

Test a parameter change without publishing it

Ask for one comparison: monthly versus quarterly rebalancing with the other inputs held fixed. Call run_model with intent=test, including for a baseline at the defaults. Save both run IDs and use compare_runs; inspect differences in inputs and windows before interpreting metrics.

A test is not adoption. If you decide to adopt tested defaults, use the separate promotion workflow. If you need different strategy logic, create a separate private model and compare its results with the reviewed model. Draft and revision tools are not currently enabled in the public MCP tool list; discover the tools available to your connection before planning a workflow. Keep the reviewed live model intact during experiments.

Maintain reviewed allocations

After review, enable the model's schedule through its settings or update_model. Use intent=refresh only when you mean to refresh the approved strategy. Inspect schedule outcomes, the latest successful run and the data date; enabled schedules can wait on data, dependencies or funding.

Ask the agent to report the latest target decision separately from current modeled weights. Broker-format files support execution by the user, not agent authority to trade. See the export workflow.

A useful completion report names the model and run, inputs changed, data dates, limitations and whether anything was published. ENSEMBLE is research software; neither an agent's explanation nor a hypothetical backtest establishes future returns.