A real research agent: give it a question and it plans, searches Wikipedia, reads pages and writes you a sourced brief — every thought and tool call streams live as it happens.
This is the real agentic loop: the LLM on Cloudflare's edge responds with a strict JSON action —
{"thought": "…", "action": {"tool": "search", "args": {…}}} — the server executes the tool
against Wikipedia's public APIs, and the observation is fed back into the model's context. Repeat until
the agent calls finish with a sourced brief, or hits the 9-step limit — where a forced
synthesis pass writes the brief from everything gathered, so you always get an answer.
The whole loop streams as server-sent events. Inference runs on Workers AI's free tier — no API keys in your browser, nothing to pay. If the model breaks the JSON protocol, the server sends an automatic repair prompt and continues.