Catch the recording 

Previously, on Dev Meetups…

We did a full walkthrough of a meeting summariser AI Agent. We demoed how Mozart Orchestrator is able to orchestrate the entire process, and how to evaluate your AI Skill prompts in Automation Anywhere.

🏗️ The Use Case: MeetingMind Task Management

At our hypothetical company, MeetingMind, decisions get made in meetings, mentioned in Slack threads, or buried in Zoom transcripts, and none of it makes it into the system that's supposed to track it. A PM ends up manually reviewing everything and updating tasks, risks, and action items after every sync.

The goal: build an agentic process that takes meeting transcripts and channel messages as inputs, reasons over them, and writes structured updates — new tasks, flagged risks, assigned action items — back into MeetingMind tracking system, with one rule: if the transcript and the channel messages contradict each other, stop and surface the conflict to a human.

It's a good candidate for an agent specifically because the inputs are unstructured and come in different formats. A transcript has speakers and timestamps. A channel message is unstructured. A deterministic process would need rigid parsing logic for both. An agent reads both and figures it out.

The stack: AI skills for transcript and channel message analysis, an AI agent for reasoning and conflict detection, Mozart Orchestrator to orchestrate the whole thing and write to MeetingMind in parallel branches, and Supabase on the backend.

🧪 Evaluating AI Skills Before You Wire Anything
The eval suite lives inside the AI skill itself. You give it a name, pick a task type (text generation for summaries and recommendations, JSON validity for structured outputs), add your test datasets with input/expected output pairs, and run. What comes back is a score per dataset, token count per run, and time taken — so you can actually compare how a prompt change affects output quality, cost, and latency before anything is live.

Shreya ran two eval sets: one testing whether the skill extracted the right attendees, action items, and risks from a transcript; another testing whether the output was valid JSON.

Pro tip: run evals on your skills before you hand them to an agent. The agent will use whatever the skill returns as reasoning context. If the skill is hallucinating attendees or producing malformed JSON, the agent's decisions will be based on that. Catching it at the skill level is much cleaner than debugging it downstream.

🤖 The Agent + Process Architecture

The write-to-platform tools — create new task, flag risk, create action item — are in the Mozart Orchestrator process, as parallel branches after the agent completes. The AI Agent accumulates its full analysis first, outputs a structured recommendation, and then the process acts on it once — cleanly, in parallel branches that don't depend on each other.

Mozart Orchestrator fetches the next unprocessed transcript (through GetNextWorkItem), passes it to the agent along with existing context from MeetingMind (current tasks, risks, conflicts), the agent analyzes and outputs its recommendations, then parallel branches in the process write new tasks, action items, risks, and flagged conflicts back to MeetingMind simultaneously. Fallback branch handles any errors with a display output form to a human.

The agent's action plan had three steps: analyze the transcript using the AI skill, fetch relevant context (existing tasks, risks, and conflicts from the MeetingMind API in parallel), then produce a structured recommendation with everything accumulated.

Watch the full build here.

Up Next...

The next dev meetup is on August 27th. Winners of our Pathfinder Hackathon will showcase their demos. Sign up here to watch them:

Want to present a use case, a half-formed idea, or anything in between? Reach out at 📧 community@automationanywhere.com