AI coding agents — Claude Code and tools like it — operate on plain-text code. That is exactly what a modelx model is. Models are saved as plain Python text, so an agent can read an entire model the same way it reads any codebase: explain what a projection does, trace how an assumption flows through it, refactor a space, extend a product, or write tests for the results. Nothing about the model is opaque to the agent, because nothing about the model is opaque to anyone.

The same applies after deployment. modelx models export as self-contained pure-Python packages, and those packages are standard Python code: AI agents and CI pipelines can import them, run them and test them directly, with no modelx-specific tooling in the loop.

Plain text is the interface

This is a practical difference, stated factually: proprietary binary model formats are not directly readable by coding agents. An agent can only work on what it can read, and what it can read is text. Actuarial models built as code get the full benefit of the current generation of AI tooling — and of whatever comes next, since plain text is the one format every generation of developer tooling has supported.

Concretely, in the modelx ecosystem:

• lifelib templates can ship CLAUDE.md project guidance so that agents start with the right context about a model’s structure and conventions.
• A modelx actuarial skill exists for Claude, packaging modelx know-how for agent use.

Governance for agent-assisted modeling

For a governance-minded audience, the right question about AI agents is not whether to allow them, but how to review their output. With modelx, the answer is the same as for any model change, because agent output is a model change:

• Git diffs show exactly what the agent changed, line by line, and changes go through the same review that any human change would.
• Dependency tracing shows how a changed formula affects results, from any value back to its precedents.
• Tests run against the model — or against its exported pure-Python package — catch regressions automatically.

These guardrails exist precisely because the models are code. A model that lives in a binary format can neither be edited by an agent nor reviewed with diffs; a model that lives in plain text can be both. That is the argument for models-as-code: the same properties that make models auditable by humans make them safely workable by agents.

Get started

pip install modelx
# or
conda install -c conda-forge modelx

Read Why modelx for the full case, or start with the tutorial.