How Micro Built an Everyday Work Assistant That Ran 1M+ Autofill Workflows in <3 Months

Micro is an all-in-one productivity platform that brings email, CRM, meeting notes, documents, and tasks into one workspace. Founded by former Google and Clearbit product manager Brett Goldstein, its public beta launch generated 2.7 million impressions.

Micro’s AI assistant connects to more than 1,000 tools to help users prepare for meetings, draft follow-ups, or build custom apps using their workspace data. It’s built on Mastra, combining agent and workflow primitives with observational memory, working memory, and Micro’s own memory system.

How Micro changes everyday work

Throughout the workday, people jump between different apps to check emails, meeting notes, documents, and tasks.

For example, when a salesperson finishes a customer conversation, they have to update the deal status, record next steps, draft a follow-up, and give teammates enough context to pick up the conversation.

Micro allows them to do it all in one app. Users can keep documents and tasks linked to the relevant deals and collaborate on them with teammates via app, slack or iPhone.

That context also supports recurring work. Skills describe repeatable work, while automations run it on a schedule in the background. For sales, Micro’s skill creates documents like buyer personas, positioning, or objections cheat sheet for sales.

Users can set an automation to refresh them regularly, they can also edit instructions for these automations. The output documents are then available when the agent helps draft an email or work out how to approach a customer.

Sales is just one area where Micro can help. The tool supports project management, content creation, reporting, email management, and startup investing.

The architecture of the Micro AI assistant

Micro uses a shared Mastra agent for chat and scheduled automations, alongside dedicated agents for specific workflows and subagents that help with tool use. The shared agent gets different context depending on where it’s running. In chat, it can ask follow-up questions. In an automation, it produces the result without starting a conversation.

Micro continuously ingests email, calendar, and meeting data from Google and other APIs to process. Mastra workflows then extract structured values, validate them against field types, and run autofill or automatic record creation.

An example can be updating CRM properties when a new email or meeting summary arrives, or labeling incoming emails to help with inbox filtering. These workflows run in the background, without users having to ask the agent to perform each update.

Micro connects people, companies, emails, meetings, documents, and tasks, so the agent can follow the relevant context across the workspace. Interconnected objects are stored in Postgres, with a custom query layer called Prism. On top of that, users can add their own documents, for example, to explain a sales process.

Micro’s memory system supports working memory for current priorities, episodic recall of past conversations, and long-term facts and preferences. These help the assistant carry relevant context across chats and automations.

To support that, Micro combines Mastra’s observational and working memory with its own system and external providers.

The agent can also work with connected tools. Micro supports custom MCP connections and integrations such as Linear so users can retrieve project information or run automations that update it.

Micro exposes its own data and actions through an API, SDKs, and CLI. Developers can use them to read workspace information and update records from their own software. For example, an engineering team can turn a standup transcript into a project document in Micro, then retrieve it through the SDK for use in a coding tool.

Helping users during live calls

Micro’s live meeting assistant uses Mastra agents and Recall.ai transcripts to help users answer questions during calls. It provides insights, talking points, and suggested questions as the conversation develops.

To keep responses fast—they achieved sub-3s speed—Micro sends the agent a rolling window of recent conversation instead of the full transcript. It streams answers as they’re generated and cancels outdated requests when new context arrives. Contact lookup and search run in parallel, with a short timeout so slower retrieval doesn’t hold up the response.

Building apps and controlling access

Users can ask Micro to build custom apps and dashboards using live data from their workspace. Micro uses a Daytona coding sandbox to build the apps, with Micro’s SDK and CLI providing access to workspace data.

Users can share apps with teammates without making them public, or publish them so anyone with the link can open them

Working with private emails and customer information makes data access the biggest engineering challenge. Micro scopes data to individual users and enforces the separation between private and shared information in code.

The team has to also make sure new collaboration features don’t reveal users’ private information.

Why Micro chose Mastra

When they started building the product, Micro evaluated a few frameworks, including LangGraph, but worried that they were too opinionated and didn’t allow much customization. So the team initially built its own agent harness using direct OpenAI API calls and Vercel’s AI SDK.

Then they heard about Mastra. Mastra’s composable primitives fit the way the team wanted to build.

“In Mastra, the paradigms that you have to follow are less strict. As a startup and at that stage, you change a lot of things and there’s always ambiguity on what the product is going to be like. So we want to have that freedom and flexibility.” — Naveen Sreekandan, Engineer

Micro could replace parts of its custom implementation as Mastra kept adding new features. For example, the team built its own support for skills first, then moved to Mastra’s native implementation when it became available.

They also appreciate the extensibility of the framework.

“When we evaluate services like Daytona for code execution, we first check how the compatibility is with Mastra and what the Mastra team recommends.” — Naveen Sreekandan, Engineer

Results

In less than three months, Micro’s CRM autofill workflow ran more than a million times, keeping records current as new emails and meeting notes arrived. Its automations also completed more than 21,000 successful runs.

Across the platform, Micro’s agents made 1.1 million tool calls in roughly six months to retrieve information and take actions for users.