Prompt to Production: Building ERP Addons With AI Agents in Ops

The Customization Bottleneck

Every business that has ever implemented an ERP has faced the same wall: the standard modules get you 70% of the way, and the last 30% — the part that makes your operation unique — requires months of custom development at enormous cost.

The traditional answer is a services team: consultants who write custom code, charge by the hour, deliver six months later, and then disappear — leaving your team unable to maintain what they built. Ops breaks this cycle entirely.

Agentic Engineering: A New Paradigm

Ops includes an embedded AI Agent that understands your business context: your data schema, your existing addons, your tenant configuration. When you describe what you need, the agent doesn't just generate code — it generates a complete, deployable addon.

Here's a real example of the interaction:

> You: "I need a VIP customer loyalty wallet. Customers earn 2% cashback on every purchase over $500. They can redeem their balance as a discount on future orders. I want a dashboard widget showing total liability."

> Ops Agent: "I'll build a `loyalty-wallet` addon. It will add a `customer_wallets` table with a balance ledger, hook into the order completion event to calculate cashback, add a redemption option to the POS checkout flow, and create a Finance dashboard widget showing total outstanding liability. Estimated completion: 3 minutes. Shall I proceed?"

The agent knows your schema. It knows which events the POS fires. It knows how dashboard widgets are registered. The output is not a code snippet — it's a signed WASM binary ready to install.

What the Agent Actually Produces

When the agent builds an addon, it generates:

• `manifest.yaml`: Declares database migrations, capabilities, event hooks, and UI registrations.
• TypeScript controllers: Business logic compiled to WASM.
• Migration files: SQL schema changes that are applied atomically on install.
• Test suite: Automated tests run against a sandboxed copy of your actual schema before deployment.
• Rollback scripts: In case something goes wrong, the addon can be cleanly removed.

The entire bundle is compiled, signed with the Ops CA, and hot-swapped into production without a server restart.

The Safety Loop

The most important question is: what happens if the generated code is wrong?

Because the addon runs in a WASM sandbox:

• Schema errors are caught during migration dry-run before any table is modified.
• Logic errors surface during the automated test run against your sandboxed schema.
• Runtime panics are caught by the Wasmtime trap handler and trigger automatic rollback.

At no point can a broken AI-generated addon corrupt your core data, expose data to other tenants, or cause downtime for other features.

From Months to Minutes: Real Business Impact

A mid-sized restaurant chain in Mexico City used Ops to build a custom "Kitchen Performance Analytics" addon — tracking per-dish preparation time, station bottlenecks, and peak hour throughput. With a traditional ERP vendor, this would have required a scoping engagement, an integration project, and a six-figure budget.

With Ops, a business analyst described the requirements in plain Spanish. The agent built, tested, and deployed the addon in 4 minutes. The chef reviewed the first dashboard within the same shift.

This is not a future capability. This is how Ops works today.

The Developer Ecosystem

For developers who want to build addons professionally — either for their own business or for the Ops Marketplace — the agentic workflow is also available via the Ops CLI:

```bash

ops agent build --prompt "subscription box inventory tracker with low-stock alerts"

ops addon test --addon ./subscription-tracker

ops addon publish --marketplace

```

The agent handles the scaffold. The developer reviews, refines, and publishes. The marketplace review process is structural (WASM capability audit), not manual code review.