Synthesis: An Autonomous Multi-Agent Research EngineSynthesis

Synthesis applies one rule to language models: don't trust output you can't trace. The product is a live view of the agent graph. A planner breaks your question into sub-questions, a pool of researchers works on each one in parallel (search, fetch, chunk, embed, retrieve), a synthesizer writes a cited report, and a critic checks every claim against its sources. All of it streams to a control room over SSE.

Agents never touch React. They emit typed RunEvents onto a many-producer, single-consumer event bus, and the SSE route drains it, so the UI shows only what the graph did. Citation checking is a separate step: the critic re-reads each claim against the passages it cites, labels confidence (supported, single-source or disputed), and can trigger a bounded revise loop. Every capability (LLM, search, embeddings, vector store) is an interface with a real and a mock implementation, selected by environment at the boundary, so the agents only see the interface. An MCP server also exposes the research tools.

Give it a question and the lanes fill in real time: tool calls landing, sources registering, a report being written with inline citations, a per-claim confidence ledger and an interactive evidence graph. At the end you get a shareable /run/<id> URL and a token and USD cost meter. Dangling citations are caught and shown, so the confidence labels come from verification and not from the model's own assessment.

“

SSE is one-way, so "stop" is a client-side AbortController and not a server signal. WebSockets would allow real mid-run control. The in-memory vector store works for demos, but pgvector is needed for anything that survives a restart, and the bounded revise loop trades cost against quality without guaranteeing either.