Open source

Agent Business Workspace

How can a coding agent work with an existing business system? Explore a design and runnable inventory example where the agent reads data and proposes changes, and the backend checks and applies them.

GitHub Documentation MIT

Give an agent a business task

“Find products with enough stock, fill in missing descriptions and prepare a weekend offer. Let me review it first.” That request spans products, inventory, content and promotion rules. The agent needs to understand how those parts fit together and what it is allowed to change.

Agent Business Workspace shares Foundax's approach: give a coding agent readable business data and instructions, let it write code to work through the task, and send its proposed changes to the existing backend for checks and execution. The design is relevant to teams adding agents to commerce, CRM, ERP or marketing software.

Three questions behind the design

How does the agent know what the data means? Explain relationships, units and rules such as available stock.
What happens when someone changes the data mid-task? Refresh the affected data while keeping useful scripts and results.
Who applies a proposed change? The backend checks permissions, business rules and the current record version before writing.

Start with the inventory example

The runnable example compares a stock count with product records and writes a list of suggested corrections. The next step checks each correction, including whether the stock covers reserved orders and whether the record has changed. If any correction is invalid, the whole batch is rejected.

With Node.js 22 or later, run these commands from the repository root:

npm run demo
npm run apply-demo

The example uses synthetic data and an in-memory service. It needs no model API key or database. The README also explains how to let your own coding agent write the calculation script. The broader design for isolated execution, cached data and automatic refresh is described in the engineering documents.

Bring a real business problem to the discussion

How should an agent interpret a CRM relationship, a warehouse unit or a campaign metric? What should happen if the data changes while a user reviews a proposal? We welcome concrete cases and designs through the repository's contribution guide and Issues.

Read the example and design on GitHub. For the interface side of the problem, see Agent Editable UI.