Public: Replit Agent coding stack

Architectural notes

Per LangChain's Breakout Agents study, Replit Agent runs on LangGraph
with a multi-agent design (manager, editor, verifier) in a ReAct-style loop and
uses LangSmith for trace observability.

Replit states Agent runs on Anthropic's Claude Sonnet: Agent 2 (Feb 2025) used
Claude 3.7 Sonnet and Agent 3 (Sept 2025) moved to Claude Sonnet 4 (normalized
here to the catalog's claude-sonnet-4-6).

Agent calls 30+ tools through a custom Python DSL rather than standard function
calling, and self-tests apps via a REPL-based verification loop with browser
automation Replit says is 3x faster and 10x cheaper than computer-use models.

Apps build, run, and deploy on Replit's own cloud/Nix sandbox (proprietary, not
catalog-mapped); automatic checkpoint commits enable point-in-time reversion.

Not publicly documented: vector retrieval, an external memory store, or routing.