OpenClaw: an autonomous markets research agent, in action

The setup

OpenClaw runs locally and speaks the Model Context Protocol. Connecting it to ClawTerminal is the same one-line registration as any other client — one endpoint, one key, ~160 tools. The difference is the leash: instead of answering a single question, OpenClaw is handed a brief and a loop, and it decides which tools to call, in what order, until the brief is satisfied. Think of it as a junior analyst who never sleeps and never fabricates — because it can only report what a tool returns.

06:00 — the sweep

The run opens with the cheapest, highest-leverage call on the whole surface: a single cross-surface recency sweep. One tool, every domain, ranked by materiality.

# OpenClaw, unattended. Illustrative run log.
> openclaw run --brief "morning markets sweep; flag what's worth a look"
[1] whats_new(since_days=1)            -> 47 events across 12 surfaces
[2] rank by materiality, drop low-signal
[3] for each survivor: resolve_entity + pull context
[4] draft note, attach sources, score vs SPY

Out of the noise, a handful of things clear the bar. A priority FDA approval prints for a large-cap pharma — the kind of single-stock catalyst worth a dedicated look. A nine-figure insider buy lands at a media holding company. The 10-year-minus-2-year Treasury spread sits modestly positive at about +0.42, with the 10-year yield near 4.45% — a quiet but real piece of macro-regime context the agent files away.

08:30 — pulling a thread

The brief says “worth a closer look,” and one theme keeps recurring across the sweep: the financing plumbing behind the AI boom. A chart making the rounds claims ordinary retirement savings are quietly funding AI data centers. Rather than react, OpenClaw does what it always does — it goes to the filings. It pulls the relevant 10-Ks and the regulator document, traces the annuity-to-GPU chain hop by hop, and separates the figures that appear in disclosures from the estimates that don’t. The verdict it drafts mirrors our own teardown of that chart: mostly true, with the scariest numbers being unverifiable, paywalled estimates rather than anything in a filing.

This is the whole point of an autonomous run. The agent didn’t need a human to tell it the chart was worth checking, and it didn’t need a human to tell it how — trace each claim to a primary source, and where there is no source, say so.

10:00 — valuing the names in the story

If the thread is AI capex, the obvious follow-up is: how are the companies at the center of it actually priced? OpenClaw pulls the fundamentals and the split-adjusted prices for the megacaps and runs the same valuation block on each — market cap, P/E, P/S, free-cash-flow yield, growth — then drafts a one-line read for each.

It flags Apple too — ~42x earnings on ~6% revenue growth, a premium for low growth carried by the buyback flywheel. None of these are calls. They are the agent putting a number and a sentence where a vibe used to be, with every figure traceable to a 10-K. A human reads four tight reads in the time it takes to make coffee, and decides what, if anything, to do.

10:20 — logging the work

Finally, OpenClaw writes its conclusion somewhere it can be graded. It posts a structured note to the ideas board, where every thesis is scored mechanically against the S&P 500 at the horizon’s end. That closes the loop: an autonomous run that produces opinions nobody ever checks is a toy; one whose track record is measured against a benchmark is a research process. The agent’s job is to be right often enough to earn the next look, and the scoreboard keeps it honest.

Why this works: retrieve, don’t recall

The reason an autonomous markets agent is even possible without it dissolving into confident nonsense comes down to one principle: it retrieves rather than recalls. A model asked from memory what was in a filing will invent a plausible answer. The same model handed a tool that returns the real filing simply reports it. Wire every claim — a filing item, a price, a margin, a leverage ratio — to a tool call, and the failure mode that makes people distrust AI for finance largely goes away. When a claim has no source, the agent says so, and the absence becomes the finding.

That is the entire bet behind connecting an agent to a real database instead of asking it to be one. You keep the model and the loop; you give it the ground truth. OpenClaw is one way to run that loop hands-off — but the same workflow runs in Claude, ChatGPT, Cursor or any MCP-capable client you prefer.

Run your own

Nothing about this run is exotic. It is one endpoint, a standing brief, and the discipline of sourcing every claim. Point your own agent — autonomous or hands-on — at the same data and the same workflow is yours. The rest of this blog is, in a sense, the output of exactly this process, written up.