How it Works

Deepwriter is a general agentic intelligence. When you provide it a prompt (and other inputs), 100s of agents immediately get to work performing 1000s of tasks over 10s to 100s of millions of tokens in planning and collaboration. Web searches can also get as high as into the 100s as it conducts "deep research". Deepwriter is simply the most powerful and flexible artificial intelligence. All throughout the process Deepwriter Agents can use the many 100s of powerful pre-built tools giving them superpower in areas where LLMs are otherwise very weak. In many cases, Deepwriter even creates its own tools and architecture on the fly. Still this is a 10,000 foot perspective of what each project undergoes.

Deepwriter's architecture consists of three main stages:

1. Planning (Optionally with a Human-in-the-Loop)

   • From big picture to details.

   • "Deep Research" can go into every Generation, regardless of type of work.

   • User-guided mode allows power users to have control over major milestones.

2. Outlining Loop

   • Outlining in Deepwriter is a recursive process that is not only conducted via text, but also via diagrammatic thinking. This is a new proprietary system invented for Deepwriter where complex relationships, sequences, state changes and more are first class citizens in the AI's thought process when words are just not powerful enough.

   • You can see the CoT and the diagrams capturing snapshots of this thinking process in the "Planning Overview" in the History or in the emails sent for each Generation.

   • These diagrams are as diverse of the outputs of Deepwriter and are a free bonus file included for each work. Some examples of diagrams from outlines below:

These diagrams are not mere summaries of text-based reasoning step, but are actual reasoning steps performed in diagrammatic form using our proprietary approach. The Agents have these diagrammatic capabilities throughout but this is especially common in the outlining process as it figures out complex relationships, sequences on its path to generating novel ideas.

Writing/Analysis/Rewriting Loop

• The final loop distills everything.

• Not only is there a loop to write/analyze and rewrite, but proprietary AI technologies that are not LLMs complete many of the writing improvements escaping the dreaded LLM-speak that is easily detected and very poor writing.

• A result of the proprietary process is that outputs don't sound like they were written by an LLM because to a large degree they were not.

• Deepwriter outputs are often detected as mostly human writing (sometimes up to 98% of a work) when competitors are typically detected at the opposite extreme (usually around 98% AI). Deepwriter makes no attempts to fool these detectors. Quite the opposite, every Deepwriter work includes a disclaimer that the work was generated by AI. In some works we have tested, this disclaimer that the work was generated by AI was the only part that is detected by AI in these detectors!

High Level Architecture: Every module in this diagram is a nested system of its own managed and evaluated in real time by an army of AI Agents.