Behind the Scenes: Building PostStream — My Multi-Agent AI Content Engine

Over the last month, I've been experimenting with agentic AI systems. Getting awar from single prompts. Introducing cooperative agents with defined roles, memory, and iteration. That work became PostStream, a lightweight multi-agent architecture I’m using to automatically plan, draft, and refine content across multiple platforms.

This post breaks down how PostStream works under the hood, why I chose a multi-agent approach, and what I’ve learned so far.

Why Multi-Agent Instead of One Giant Prompt?

Traditional single-prompt generation works for simple tasks, but it quickly breaks down when you need:

• consistent tone
• platform-specific formatting
• reliability
• structure
• reasoning + creativity
• guardrails

Trying to make one model do everything — plan → draft → revise → format — leads to:

• unpredictable outputs
• weak structure
• repeated content
• fragile prompts that break easily

A multi-agent architecture solves this by giving each agent a single responsibility.

PostStream breaks content creation into a pipeline:

Each step is smaller, easier to tune, and easier to debug.

The PostStream Architecture

Planner Agent

The planner defines the high-level structure:

• target platform (X, Reddit, LinkedIn)
• tone (casual, editorial, energetic, technical)
• required sections
• how long the post should be
• the core angle or hook

It outputs a structured JSON plan — not prose.

Writer Agent

The writer converts the plan into a first draft.

This draft is intentionally loose:

• focuses on ideas
• gets the content onto the page
• doesn’t worry about polish
• doesn’t enforce tone too early

Keeping the writer “creative-first” helps avoid the rigidity that often comes from early editing.

Editor Agent

The editor checks the draft against tone rules:

• clarity
• readability
• sentence length
• voice consistency
• trimming redundancy

It also enforces platform conventions. For example: X prefers conciseness, Reddit allows longer exposition, LinkedIn benefits from structured sections. If the post breaks these rules then it is sent to the revision agent.

Revision Agent

The revision agent acts as the 'enforcer.'

It revises the content based on the feedback from the editor. Correcting issues with:

• banned phrases
• formatting issues
• JSON schema mismatches
• missing sections
• platform-specific hazards
• potential Reddit automod triggers
• X character limits
• markdown correctness

The revised version is then sent back to the Editor Agent to be checked again. The flow will retry this Editor / Revison loop until a clean version is verified by the Editor.

Why This Feels More Reliable

After working with PostStream for a few weeks, several things became clear:

1. Smaller agents are easier to reason about
   If the Planner fails, I know exactly where to look.
   If the Editor drifts in tone, I update one file.

2. Creativity and structure need to be separated
   The Writer should be free to generate ideas.
   The Editor should make them coherent.
   Trying to do both at once produces the worst of both.

3. Tone belongs late in the pipeline
   Injecting tone during the first draft makes posts bland and same-sounding.

4. Platform-specific constraints matter
   Reddit automod has nothing in common with X’s character limits.
   Each platform needs its own revision rules.

5. Multi-agent loops build trust
   I get more predictable, consistent results because each agent has a defined contract and responsibility.

What’s Next for PostStream

I’m currently expanding the system with:

• subreddit-specific automod rule checks
• analytics-driven rewrites (learn from what performs well)
• thread generation for X
• multi-level memory across agents
• stronger error recovery on malformed JSON
• a higher-level supervisor agent to direct multiple content tasks
• scheduling and batching routines

Long-term, I want PostStream to be a reliable content partner — something that can generate weekly updates, cross-platform variants, and structured posts without creating additional overhead.

Closing Thoughts

Multi-agent systems aren’t magic, but they bring a surprising amount of reliability and clarity to tasks that normally feel unpredictable with single-prompt LLM use.

If you'd like to follow more of these experiments (including PostStream and my Domain Discovery Engine work), you can always check new posts here on 3EF Studio.