Landing Page
Calendar Page
Live Updates
Agent Overview

MediaFlow

Tagline: Buffer for autonomous brand agents. You hire AI agents, not write posts.

Project Description

The Inspiration

Traditional social media marketing is broken. Human agencies are too slow and expensive to scale, while current AI tools rely on single-prompt generators that produce repetitive, low-engagement "AI fatigue" content. We realized that small businesses and creators spend significant, expensive time on social media management. We asked ourselves: What if instead of using an AI tool to write a post, you could just hire a fully autonomous AI marketing team?

What it does

MediaFlow is a multi-agent marketing automation platform. A business describes itself in natural language, and a team of AI agents autonomously generates a complete marketing strategy, creates platform-optimized content (video, carousels, images), quality-checks every piece, and schedules publishing across social platforms.

You aren't just chaining prompts; you are deploying a virtual "Marketing Director" that manages a "Strategist," a "Critic," and specialized media creators.

Main Hackathon Track: Flicker to Flow (Productivity)

MediaFlow removes the massive professional toil of social media management. By deploying autonomous AI agents that self-improve, businesses can reclaim hours of productivity every week while maintaining an agency-quality brand presence.

How we address the Company Challenges

🏆 Fetch.ai - Agentverse & ASI:One (Primary Track)

MediaFlow is deeply integrated with the Fetch.ai ecosystem to develop a single or multi-agent orchestration that demonstrates reasoning, tool execution, and solves a real-world problem. We built a true multi-agent system where agents are genuinely autonomous.

ASI:One Integration: While we experimented with Claude for early extraction, ASI:One is the exclusive LLM running at demo time. It powers our brand extraction, marketing analysis, slate generation, and our critical 5-axis scoring rubric.
uAgents Framework: Every agent (Agent, Context, Protocol) is built using Fetch.ai's uAgents.
Real Autonomous Communication: Our Head Agent dispatches tasks to the Strategist, Critic, and Publisher via ctx.send(). Cognition agents wake on a cadence and run the full pipeline without human intervention.
Agentverse Discoverability: All agents are registered on Agentverse and implement the mandatory Chat Protocol.

🏆 Cognition - Augment the Agent

We successfully built an integration that makes AI agents measurably more capable, or removes the friction and toil they can't yet handle on their own.

The Critic Agent: Standard AI agents just rubber-stamp generation. Our Critic enforces a 5-axis rubric and must reject at least one piece of weak content per slate, proving that our agents can self-correct.
Publisher Reliability Subsystem: Built end-to-end by Devin, this subsystem handles idempotency and dead-letter queues to ensure agents can publish reliably at scale.
Performance Harvesting: A daily cron agent pulls data from per-platform native APIs and feeds engagement data back into the Strategist's context so the agents learn what works over time.

🏆 Happy Hacking Keyboard - Type Beyond

We made sure the keyboard is central to the experience of commanding our AI agents within our web application. Managing a multi-agent system usually requires clicking through clunky dashboards, but we transformed keystrokes into something more expressive, interactive, and meaningful for a browser-based workflow.

Browser-Safe Custom Keybinds: We carefully designed custom keybinds that explicitly do not interfere with standard web browser shortcuts. This allows users to rely heavily on muscle memory without accidentally triggering native browser actions like opening new tabs or refreshing the page.
Keyboard-Driven Orchestration: Users can seamlessly control the live streaming of agent activity directly from the web app. While the Head Agent and Critic are actively debating content in the background, a user can pause the pipeline, approve slots, or instantly inject text feedback on the fly through intuitive keystrokes—all without ever needing to touch a trackpad.

How we built it

Backend: Python-based agents managed within a uv workspace using the Fetch.ai uAgents framework. We utilized a Bureau pattern to register our Strategist, Critic, Publisher, and dynamic cognition agents.
Models: ASI:One handles all natural language reasoning, strategy, and extraction at runtime. Google Veo API is used for platform-specific video generation.
Media & Publishing: A custom Pillow-based renderer for image carousels, followed by direct platform APIs (LinkedIn, X, Instagram, TikTok, YouTube, Bluesky) for cross-platform distribution.
Frontend & Database: Next.js 16 web application (React 19, Tailwind) connected to Supabase (PostgreSQL with Row Level Security for tenant isolation).

Challenges we ran into

Building a system where agents don't just talk, but actually enforce quality, was incredibly difficult. Initially, the LLMs would just approve every generated post. We had to heavily engineer the Critic agent to be adversarial, forcing it to reject content that scored below a 3.5/5.0 threshold.

Accomplishments that we're proud of

We successfully built the "hire an agent" paradigm. Each brand on our platform gets its own persistent, autonomous uAgent on Agentverse with a stable address. We are also incredibly proud of the fact that our multi-agent loop is load-bearing—our Critic actually rejects bad content, meaning the system genuinely thinks and revises rather than just executing blindly.

What we learned

We learned that orchestrating multiple agents requires strict data contracts. Standardizing inter-agent communication using an AgentEnvelope JSON wrapper was the breakthrough that allowed our Strategist, Critic, and Design Director to talk to each other without hallucinations or data loss.

What's next for MediaFlow

We plan to introduce agent-to-agent payment protocols so our Worker agents can charge micro-transactions for highly specialized asset generation. We also plan to expand our MCP (Model Context Protocol) Server capabilities to allow massive enterprise systems to hook directly into our autonomous marketing bureau.