Vesti — a multi-agent research desk for retail micro-cap investors.

A retail investor hunting micro-caps is outgunned by the search space. Five thousand tickers, a handful of hours a week, no analyst team, no screener discipline that survives contact with a compelling story. The work that actually matters — reading fundamentals against the chart against the news against the risk profile — collapses into whichever signal was loudest that afternoon. Conviction forms on half the evidence or doesn't form at all, and the journal, if one exists, records outcomes without recording the reasoning that produced them. What dies isn't the return; it's the feedback loop that would have made the next decision better.

A LangGraph workflow runs seven agents against the ticker universe on a schedule. The market scanner narrows five thousand candidates to a shortlist against explicit price, market-cap, volume, and beta bands. Four analysts — fundamental, technical, sentiment, news — score each candidate in parallel with their own tools and their own Claude prompts. A risk manager inverts the composite. An orchestrator applies a weighted blend (technical 30, catalyst 25, fundamental 20, sentiment 15, risk 10) and writes a top-five recommendation set with a synthesised thesis for each name.

The investor reads the thesis on the recommendations surface, sees the score breakdown behind it, and records a decision in the trade journal — act, skip, or defer. Each agent run is checkpointed against Postgres through LangGraph's PostgresSaver; every score, every reasoning trace, every rejected candidate is preserved in agent_runs and recommendations. The delta between what the orchestrator proposed and what the investor booked becomes the artefact the next review cycle is measured against.

• skills over controllers — each agent (market_scanner, fundamental_analyst, technical_analyst, sentiment_analyst, news_monitor, risk_manager, orchestrator) is an autonomous unit with its own prompt, tool surface, and scoring contract. The workflow composes them; it does not dispatch to methods on a service.
• harness over model — LangGraph owns the state graph, the checkpointing, the cancellation semantics, and the per-node logging. Claude reasons inside a node; it does not drive control flow. Swapping the model does not touch the orchestration.
• events over CRUD — agent_runs, recommendations, trades, and portfolio_snapshots are append-only. The portfolio row is a projection; the journal is the system of record. Historical theses remain queryable against the decisions they produced.
• feedback as interaction — the accept/skip/defer decision on a recommendation is the signal. The trade journal is not downstream reporting; it is the input the next scoring cycle is calibrated against.