Weave project url: https://wandb.ai/sinloes-moteo/webflow-si/
Orchestrial is a self-learning web agent. It optimizes its navigation via an adaptive flow chart to learn from its mistakes. It collects host-specific considerations on scraping which inform its orchestrating LLM, that both specifically minimize redundant and slow actions, and improve on failure.
It lets research, growth, and operations teams launch a goal-driven browser agent that navigates complex sites, learns from failures in real time, and returns structured intelligence plus replayable browser sessions. The project is useful for automating competitive research, catalog enrichment, and multi-step QA flows that normally require analysts in the loop.
The orchestrator snapshots the live DOM and run history, hands that context to the LLM agent orchestrator (sonnet-4.5), receives a Stagehand-compatible plan (selectors + natural-language actions), and executes it through Browserbase. After every attempt the loop updates failure statistics, host-specific heuristics, and an evolving flow chart that reflects the agent's hypotheses.
Dashboard → orchestrator /runs → Stagehand browser session → extraction result → repository/report generator → SSE timeline + React Flow visualization. Host tool insights, repaired data, and session URLs feed back into Claude on the next turn so retries adapt instead of repeating brittle selectors.
Weave: logging outputs and agent behavior to feed into post-run evaluator model
Log in or sign up for Devpost to join the conversation.