Workplace agents

People already work in Jira, Google Calendar, Gmail and PaddyField, and the assistant is only useful if it can act in them.

I built one specialized agent for each tool:

• Jira: create and look up tickets.
• Calendar: find free time and book meetings.
• Gmail: draft approval emails and other messages.
• PaddyField: log time.

Each agent is isolated, with a small context of its own. This is the same reason as the orchestration: I don't want the context to grow, and isolation is the key. The agents can talk to one another (the Jira agent could ask the Calendar agent for something), but I chose not to build it that way, because it adds cost to the system in latency, complexity and design.

Access is the user's own. The user first signs in to the desktop application and connects their applications through the OAuth flow in the app. The tokens are stored in a secure place, along with the permissions the user granted, and the agents act on the user's behalf within them. Before every agent call or tool call, the app checks that the person has the access it needs. If they don't, it stops there. The same check keeps restricted agents, such as HR, to the roles that are allowed to use them.

None. At that time there were no ready-made integrations such as MCP servers, so we built our own AI tools and skills for each system.

Developed multiple specialized agents for interacting with workplace applications and internal systems, including:

• Jira
• Google Calendar
• Gmail
• PaddyField

Used LangChain and supporting AI tooling to enable agents to perform domain-specific workplace tasks and participate within the broader multi-agent workflow.

This allowed Auggy to provide a unified assistant experience across multiple workplace systems.

• Maintenance. Whenever one of those APIs changes, we have to update the integration. There was no choice then, so we managed and maintained every integration by hand. These days it's much simpler with MCP.
• Agents don't talk to each other. Keeping them isolated means each task stays in its own tool, and anything that needs two tools goes through the orchestration.

The assistant acts inside the tools people already use, with each agent focused on one job.