Atlas — Manual + Auto-Research: an open-source system design playground
From internal tool to open source: system design, storyboards, and stock research powered by Claude + Perplexity and an autonomous agent.
What it is
Atlas is a dual-mode application: Manual Mode (select a component → prompt → refine → save) and Auto-Research Mode (define a research brief → autonomous agent loops → human review → finalize). The app is built for long-horizon research workflows — think multi-document analysis, codebase reasoning, film storyboarding, and deep stock due diligence.
How it works (high level)
User defines goal & success criteria.
Agent builds context from the 1M token window, fetches research (Perplexity), runs analysis (Claude), generates component drafts, and saves versions.
The human reviews flagged uncertainties, edits, accepts final components, and exports (MD, PDF, blog).
Why now?
Two recent trends make this exciting:
Anthropic’s broader 1M-token context support (so agents can hold massively more context while working) 1M token, and
Karpathy’s autoresearch pattern that demonstrates autonomous, auditable experiment loops. https://github.com/karpathy/autoresearch
Open source + license
This repo will be released under MIT. Expect: architecture diagrams, the AI orchestration service, context window helpers, RAG & caching patterns, agent loop examples, and export pipelines.
How you can help / contribute
Star/watch the repo when it goes live
Try the demo & file issues for use cases you need
Contribute connectors (other LLMs, cloud storage, export formats)
TL;DR
This tool turns long, messy research into an auditable, iterative loop where humans are always in control of decisions. Early release to contributors in 2–3 weeks; public repo launch shortly thereafter.