One agent runs the entire loop: it reads the evidence, debates every view across model families, simulates how different market participants will react — and writes the book.
Use cases
What is actually driving markets right now — assembled from the evidence and the causal graph, not limited to headlines.
Helps traders surface what the market may have wrong — potential mispricings, and the overlooked lines of reasoning behind them. For policymakers it works as a think tank: the same evidence base and simulation, pointed at a policy question.
Ask across the whole evidence base at once — filings, flows, positioning, macro vintages — and get an answer with its reasoning and citations shown.
Simulates the different market participants in each market — and the transmission mechanics between them — to preview how an event propagates before it trades.
Why an agent
A vast body of official macro data and private indicators, read continuously, to one standard — every reading sourced and filed into one ledger. Human coverage has to choose; the agent does not.
An independent agent takes the seat of every market participant — reasoning from that participant's own objectives, its constraints and the mechanics of how it trades — and rehearses the market's reaction before any position exists.
Humans — and single AI models — are unreliable narrators of their own assumptions and beliefs; the answers come out distorted. The agent's discipline is structural: every thought is recorded, every claim names what it rests on, and a second model family is tasked with attacking the conclusion.
contact@globalmacroagent.com
The engine
It first reconstructs what the market believes — the consensus narrative, what is already priced, and who is positioned where.
Then it derives its own view from primary data and facts alone — vintage-correct evidence, code-computed measurements — deliberately apart from the consensus it just mapped.
Finally it simulates the different market participants and the transmission between them, to predict how the market digests the gap between consensus and evidence.
Coverage
The method is breadth without selection: a vast body of official macro data and private indicators, read continuously to one standard — every reading sourced, time-stamped, and filed into a single evidence ledger. A human desk has to choose what to read; the agent does not.
The model
The engine follows how information transmits across families of evidence, and it keeps facts strictly apart from assumptions. The analysis is driven by the relationships between the data, not the data itself — several coherent worlds held at once, weighted and retired by evidence.
Simulation
Each market is argued by independent agents, one standing in the seat of each participant class — reasoning from that participant’s own objectives, its constraints, and the mechanics of how it participates — to simulate what it does next. A central adjudicator then weighs the long and short forces, and their weight in the market, into a single ruling on who sets the price.
Honesty
Every statement lives in a three-part ledger — fact, inference, assumption — and a fact must carry its source. Claims declare what they rest on, so when an upstream fact is overturned, the contamination is computed and everything built on it is revoked, in cascade. Tensions close only by ruling; “I don’t know” is a legal answer.
Process
Every round compresses a vast body of analysis, layer by layer, into a handful of conclusions — and every conclusion must meet the market that follows: each states, in advance, what would prove it wrong, and the forward market does the grading.
The full breadth of data and indicators becomes a ledger of claims — every one a sourced fact, or an inference that names what it rests on. Facts and assumptions never mix.
Bull argues bear across separately trained model families, and every simulated participant states what it would do from its own seat. Claims that cannot survive the argument die here.
The central adjudicator weighs the long and short forces, and their weight in the market, into one ruling per market. A contested ruling shrinks the size it is allowed to carry.
What remains is a handful of volatility-budgeted positions, each shipped with the condition that would prove it wrong. The forward market does the grading — no falsifier, no trade.
Philosophy
The agent’s reasoning distils the philosophy of the investors who defined global macro — Soros’s reflexivity, Bridgewater’s radical truth-seeking. A market view acts on the world, and the changed world feeds back into the view; the discipline is to reason inside that loop without fooling itself.
Limits
The same ledger that records the agent’s views records its boundaries.
run 31598 · 2026-09-05
It starts with what the market believes: the narratives moving markets right now, each held as a competing hypothesis and scored for strength and crowding — none forced to converge; the data retires them.
Treasury issuance, the federal interest burden, and the changing composition of the marginal buyer of US duration.
Geopolitical energy-supply risk as a transmission channel into headline inflation and bond yields.
Sticky services inflation and re-accelerating price gauges pressing against the FOMC’s policy-rate path.
Internal FOMC dispersion and labour-market cooling read as constraints on how far the rate path can rise.
in the headlines this round: global bond sell-off deepens on US–Iran escalation · August jobs report and Fed hike bets · stocks rise as hike odds ease on Waller · yen surges on BOJ rate-hike and intervention bets
Adversarial cross-examination
The agent cross-examines its own causal graph against the market’s narratives, tension by tension. What survives is a short list of potential mispricings and overlooked lines of reasoning — each with a diagnosed cause and a written catalyst.
The strip prices contained rates vol and no funding-risk premium while reserves drain with the RRP buffer exhausted and the ample-reserves floor unquantified. Diagnosed cause: attention scarcity — no market narrative carries the mechanism at all. The written catalyst: a quarter-end or tax-date SOFR print materially above IORB with standing-repo usage jumping.
Despite record debt, the widest monthly deficit on record and foreign custody down to $2.60tn, duration keeps clearing — the custody decline is a composition shift in the buyer base, not a shrinking total bid. The crowded absorption-failure trade is the error; the true fragility sits elsewhere.
The committee delivers the hike and holds a restrictive stance while the chair–committee circuit holds; the view that internal dissent restrains the path fails, and the twelve-month priced change stays hike-side.
Even with WTI above $90 and gasoline above $4, market-based inflation compensation stays anchored — positioning for an oil-driven breakeven blowout loses while the fiscal and supply channel keeps carrying the sell-off.
the written single point of failure: all three views transmit through one reaction function — the FOMC under its current chairmanship; a change in that chain unwinds all three at once · falsifiers on file: a durable 10-year below 4.25% with no fiscal event and no runoff halt kills the supply view; delivered hikes kill the policy view
every tension between the agent’s graph and the market’s is closed by a written ruling — it audits itself as hard as it audits the market, and files every correction the same round.
the epistemic ledger
A few lines from this round’s ledger, essentially as filed: every claim is a sourced fact or an inference that names what it rests on — and where the evidence ends, the ledger says so explicitly.
Foreign official custody at the Fed fell to $2.60tn while total foreign-held federal debt rose to a record $9.27tn — private buyers, via UK and Cayman conduits, absorbed the increase, while MOF weekly data show Japanese repatriation at the margin.
Shelter CPI lags market rents by 12–18 months, so the near-term core contribution is governed by the pipeline of past rent growth and the current shelter rate — not by the index sitting at a five-year high, which is trivially true of any component with positive cumulative inflation.
Dealer take-up was 1.85% on the 30-year and 6.84% on the 10-year, against 25–33% on twos and bills: end-investors cleared the long end while dealers backstopped the front. Absorption measured auction by auction, not narrated.
and a claim can be a refusal: where no published figure exists yet, the ledger writes “unknown” — never a plausible-sounding guess.
＋ The long arc · slow variables near their turning conditions
Multi-agent simulation
Every market is argued by the simulated participant classes — each files a stance, a size and its reason. A sample of this round’s filings, then the desk that decides who actually sets the price:
US TREASURIES · DURATION
US DOLLAR · CURRENCY
US EQUITIES · INDEX
GOLD
⚖ The adjudication desk · hard evidence only — positioning motion, event digestion, forced flows; majority opinion never sets the price, the marginal moving participant does
The recommended book · run 31598
Five markets. Each direction is a settled argument — the reasoning ships with the position.
weights are % of capital · volatility-budgeted, 10% annualised across five markets · every stance carries its falsifier · paper-traded, marked daily