MAgent - Autonomous Content Marketer

• 

  Flowchart

• 

  Trigger Inputs from the user (optional - for demo purpose)

• 

  Service Logs

About the Project

Inspiration

We wanted to build a marketing tool to help indie hackers and early startup founders who must juggle multiple roles—from development to marketing - so they can focus more on building great products. Our experience as an indie hacker and conversations with early-stage startup founders underscored the challenge of maintaining social media consistency. This led us to ponder: "How might we help indie hackers and early-stage founders automate their content marketing pipeline?"

What it Does

MAgent is an agent that spontaneously plans marketing campaigns and updates its plans based on external triggers like cron jobs. The trigger currently tells the agent to check the latest news articles from the past 24h and generate content if any of the articles are relevant to the company's target customer. If the agent finds relevant articles, it drafts and iterates on the output by evaluating the quality based on user-defined criteria. Finally, it posts the draft on Slack for confirmation.

How We Built It

We wanted the agent to communicate with the team, so we built the interactions on Slack, enabling all management tasks on the app. The agent works on express server and can be triggered manually by slack messages. When the agent tries to accomplish a task, it follows 4 steps: plan the action, build hypothesis of the output, execute the action, and evaluate if the output is aligned with the hypothesis. By dividing the steps like a PDCA cycle, we could force the agent to complete the evaluation step-by-step. It also helped us evaluate its output of each step to debug any problem.

Challenges We Faced

Our agent got stuck in the loop when it failed to fetch the latest news articles because it couldn't identify the problem. We are fetching news from an external API and had a bug fetching the articles. The agent repeatedly tried calling the API with different keywords because it assumed the problem was the query the agent generated to search for the articles.

Accomplishments We're Proud Of

We built an agent reflecting on its actions and posting its thoughts on Slack in a few hours.

What We Learned

Using an agent to evaluate its action requires more work because the agent can't accurately identify the problem. (It's too probabilistic) One idea we have is to build some root cause analysis tool that the agent can call to help systematically identify the problem to plan its next action, but that is out of our scope.

What's Next for MAgent - Autonomous Content Marketer

In addition to general improvements in the output quality, we see three possible upgrades. 1. Adding more data sources like scientific publications and web searches, as requested by the agent during the reflection on the output. 2. Letting the agent post and get feedback using the number of views and conversion. 3. Broadening the type of output it can generate.

Built With

• express.js
• groq
• javascript
• llama
• openai
• replit
• slack
• supabase