ECHO: Multi-Agent Meeting Workflow Autopilot System

ECHO is a multi-agent automated workflow system for enterprise meeting scenarios, developed for the 2026 Milan AI Week AI Agent Olympics Hackathon, specifically designed to address the post-meeting work black hole. The system automatically joins Zoom, Google Meet, or Teams meetings via Recall.ai, and after the meeting, five specialized agents collaborate to complete follow-up tasks such as CRM updates, follow-up email drafting, and task creation. Its core innovation lies in the 'auditable agent memory'—each agent operation can be traced back to the meeting audio clip that triggered it and the agent discussion process, addressing the trust barrier for enterprises deploying AI agents. ECHO's goal is to achieve 'When you leave the meeting, the work is already done.'

For knowledge workers, the end of a meeting often means the start of another round of trivial work: organizing minutes, extracting action items, updating CRM, drafting emails, etc. According to statistics, managers spend an average of 5 hours per week on meeting-related follow-up work. Existing meeting tools (such as Otter, Fellow) only generate meeting minutes and do not solve the execution problem. The ECHO project was born to fill this gap, aiming to become a complete multi-agent workflow autopilot system.

ECHO adopts a specialized multi-agent division of labor model, coordinating work via a message bus:

1. Action Extraction Agent: Uses Featherless domain-specific model to extract commitments, tasks, and action items from meeting transcript text;
2. Stakeholder Classification Agent: Uses speaker separation technology to identify participants and their roles, laying the foundation for subsequent CRM updates and email routing;
3. Decision Agent: Performs complex reasoning based on the Gemini Pro model to determine operation priorities and execution order;
4. Communication Drafting Agent: Generates personalized follow-up emails that comply with enterprise norms;
5. Execution Agent: Calls external APIs such as HubSpot, Gmail, and Linear to convert decisions into concrete operations.

ECHO is deployed on a Vultr Tokyo region virtual machine, with the frontend built using Next.js and the backend using PostgreSQL with the pgvector extension to store vectorized data. The data flow process is as follows:

1. User connects Google Calendar;
2. Recall.ai automatically joins the meeting, and audio is pushed to the ECHO backend via Webhook;
3. Speechmatics completes batch transcription and speaker separation, generating timestamped text;
4. Inngest workflow engine schedules the five agents to perform tasks;
5. Operation records are linked to meeting audio clips (stored in Vultr Object Storage), supporting auditing and vector retrieval.

The biggest barrier for enterprises to deploy AI agents is trust. ECHO solves this problem through 'auditable agent memory': each agent operation is linked to the 30-second meeting audio clip that triggered it and the agent discussion process, allowing users to view the decision-making basis. In addition, the semantic search function supported by pgvector can quickly locate historical meeting discussions, making ECHO a long-term repository for enterprise meeting knowledge.

ECHO deeply integrates the hackathon sponsor technology stack: Recall.ai provides meeting capture capabilities, Speechmatics supports transcription and speaker separation, Gemini models handle reasoning and text generation, and Featherless provides domain-specific models. These technologies are not decorative; they are key components that form the core functions of the system, reflecting the product's technical maturity.

ECHO's business value is quantifiable: based on saving 5 hours per week at $50 per hour, each manager can save approximately $13,000 per year. Compared with existing tools, ECHO's core difference lies in the leap from 'information summary' to 'action execution'—existing tools end after generating minutes, while ECHO continues to complete follow-up work. The improvement in CRM data quality directly affects the accuracy of sales forecasts, providing enterprises with a clear way to calculate ROI.

ECHO is open-sourced under the MIT license, allowing developers to audit code, contribute improvements, and adapt to new tool integrations. The project's quick start guide shows that it takes about 70 minutes from cloning the code to running the product, lowering the adoption threshold. Future directions include: supporting more meeting platforms (such as Tencent Meeting, DingTalk) and enterprise tools (such as Salesforce, Jira), real-time meeting assistance, multi-language support, upgrading agent collaboration models, and extending auditable agent memory to areas such as code review and customer service.