What is an agent-built CRM?

The definition, and what it is not

An agent-built CRM is a CRM application that a coding agent — Claude Code, Codex, Gemini CLI or similar — assembles as source code from a framework's contracts, under a human's review, in the human's repository. To build your own CRM with a coding agent is to state a business goal to the agent and receive checked-in modules, workflows and policies you can diff, test and keep. The defining property is where the work lands: in source the operator owns, not in a vendor's configuration store.

The term is easy to confuse with two things it is not. It is not an AI CRM — a product whose runtime features are model-driven — and it is not an autonomous salesperson: in an agent-built CRM the model's authority ends when the source is merged. The application that results is deterministic software whose commercial state changes pass through explicit services and workflows, whatever runtime AI its owner later chooses to add.

How Accordo supports the build

The agent needs to know what already exists before it writes anything, so one command reports the composed application — packages, capabilities, resources, actions, policies, providers — read from checked-in source, never from a live process (C-14). A Solution Plan is a checked-in file bound to that inspection's fingerprint, so a plan written against an application that has since moved reports itself stale rather than silently misleading the next session; nothing executes it (C-15).

Writing goes through the same declared surfaces a human would use: module manifests that generate migrations, services, REST resources, SDK methods and Admin screens (C-01), and package scaffolds for whole domains. The local MCP server exposes project context and narrow tools, and anything that generates code or destroys state is dry-run unless an explicit apply flag is passed (C-18) — the agent proposes, the diff decides.

What the rails do not include is a measurement. The framework publishes a build benchmark protocol and has not run it, so no rate of successful agent builds exists, and this page quotes none (L-03). What is checkable instead is narrower: the commands above exist, their contracts are tested, and the verification gate runs the whole suite on every push (C-20).

What the built application is, and where it runs

The output of an agent build here is a local application: Node 22, SQLite persistence, no authentication, tenancy or RBAC, so it runs on one developer's machine and must not be exposed to a network (L-01, L-02). It is a working system to develop against and reason about, not a system a team signs in to.

Ownership follows the same mechanics as any Accordo project: the framework arrives by scaffolding from a checkout, nothing installs from npm, and upgrading means merging — the reserved registry name is still an empty placeholder (L-08). The agent that built the application does not remain attached to it; what remains is source, tests and the evidence trail the framework wrote while the application ran.