Problem
People spend enormous effort building custom AI agents from scratch. But the agents already exist: Notion AI, Claude, ChatGPT, Codex. The actual bottleneck is not agent capability but the system they operate in (missing context, unclear decision boundaries, no checkpoints). Fully autonomous agents running in a black box produce unreliable results.
Solution
Notion workspace built on the PARA principle (Projects, Areas, Resources, Archive) with an AI agent as the orchestration layer. Workflows decomposed into small, reusable AI Skills with human-in-the-loop checkpoints at the right places. Persistent project memory across sessions, structured templates for consistent project setup, and a central skills hub for recurring tasks.
Result
Single system for all projects, from academic research to software development. The AI agent operates with full workspace context instead of per-session amnesia. Reliable automation through decomposition: the agent proposes, the human confirms, the agent executes.
Lessons Learned
- The agents already exist. The bottleneck is not capability but the system they operate in: what context they have, what decisions they can make autonomously, where a human needs to step in.
- Decomposing workflows into small, human-in-the-loop steps is what makes AI automation reliable. A skill that runs five steps with a confirmation point after step three produces better results than a fully autonomous agent that runs all ten steps unsupervised.
- Persistent project memory across sessions changed everything. Without it, every conversation starts from zero. With it, the agent picks up where the last session left off with full project context.
- The system is model-agnostic by design. The same workspace structure, skills, and project memory work whether the orchestration layer is Notion AI, Claude, or GPT. Switching models changes the output characteristics but not the workflow.
Deep Dive
The system uses Notion’s database architecture as the foundation. Each project has a structured page (Meta, Current Focus, Description, Work Areas, Artifacts, PM & Changelog), a persistent Project Memory (decisions, architecture, open questions), and a Changelog that captures every session’s output. Reusable AI Skills on a central hub page define workflows for recurring tasks: project initialization, content writing, code handoffs, maintenance routines. Each skill has a clear trigger, workflow steps, constraints, and output format. The human-in-the-loop pattern is built into the skill design: the agent proposes, the user confirms, the agent executes. This avoids the black-box problem of autonomous agents making cascading decisions without checkpoints.