Excited about launching Rakenne here on Uneed!

The Problem I noticed: Most document agents fail because they are too "open-ended." Give an LLM a complex compliance report task, and it eventually wanders off-script or misses a subtle expert requirement. Massive system prompts don't scale when you have 50 different templates.

One solution I built is Rakenne (Launching today!): A SaaS that treats Markdown as the source of truth for agentic logic. Instead of a 2,000-word prompt, domain experts (lawyers, engineers, auditors) define the "workflow" in Markdown. The agent then runs this server-side via RPC, using the Markdown structure to maintain state and conduct a guided dialogue with the user.

How the "Markdown Logic" works:

Structure = State: Each Heading is a document section (or "step" in the state machine).

Instruction Blocks: We use standard Markdown syntax (like blockquotes or custom tags) to define specific "Agent Tasks" for that section.

User Inputs: Defined as variables within the text that the agent is responsible for "filling" through conversation.

The Tech Stack:

Engine: Built on the pi-mono coding agent (RPC mode) for server-side execution.

Frontend: Lit web components for a modular, fast UI that handles the agent-user dialogue.

Architecture: Multi-tenant isolation ensures expert workflows don't bleed across accounts.

Why Markdown? It’s the ultimate "low-code" for domain experts. They don't need to know Python or JSON schema—if they can write a structured document, they can now program an agent.

Looking for feedback on: Does this "constrained" approach to agents make sense for professional docs, or is the industry moving toward fully autonomous (and unpredictable) agents?

Ricardo

Activity

Products

Recently upvoted

Posts