Yala: multi-agent workspaces inside Google Sheets

Context

Spreadsheets are where most business analysis actually happens, and they are also where most people meet the limit of what they can automate. Yala started as a personal R&D playground for LLMs, AI workflows and code automation, built end to end as a continuous Vibe Coding experiment.

Problem

Connecting a language model to a spreadsheet is easy; making it useful is not. A single prompt box inside a sheet produces one-off answers. What the work needs is a way to express a workflow: several steps, each with its own instructions, running over structured data that already lives in the grid.

Approach

Yala uses the spreadsheet's own structure as the orchestration model: tabs, rows and cells become agents and automated tasks. Built on Google Apps Script and connected to the OpenAI GPT-4o API, it supports multiple customisable agents, dynamic variables drawn from the sheet's own columns and full interoperability with native Sheets functions, so an AI step and a formula sit in the same pipeline. I conceived and built it single-handedly, directing AI agents through discovery, modular architecture, code and UI generation, the pricing plans and the Stripe integration, Google Workspace Marketplace compliance and every piece of launch content.

Non-obvious decisions

• Use the sheet as the interface instead of building one beside it. A tab is a namespace, a row is a task and a cell is an argument. The user already knows the model, so there is nothing to learn.
• Ship pricing and Marketplace compliance as part of the product, not after it. A Workspace add-on that cannot be listed is a demo, so the plan structure, the Stripe integration and the Marketplace review were designed alongside the feature set.
• Keep it a living prototype on purpose. Yala exists to be iterated on and learned from, which is a different objective from a product that has to be stabilised, and pretending otherwise would distort every decision inside it.

Outcome

A working Google Workspace add-on that turns a spreadsheet into a multi-agent workspace, and the R&D surface where I test LLM orchestration, pricing and go-to-market decisions end to end, alone.