Blog · 2026-10-26 · Ashutosh Agarwal

Introducing Matterfact Agent Swarms

The frontier of AI in finance: autonomous research agents that build an investment case as a team and keep score. See what swarms found on Meta, Hims & Hers and Borr Drilling.

On 1 September we gave a swarm of AI agents one job: work out what Meta is worth.

Over the next seven weeks, the swarm:

• flagged about $356B of Meta data-center leases signed but not yet started, a bill that became one of the two biggest items when it revalued Meta after its results;
• kept a New Mexico court case on its list after nearly dismissing it twice; that case ended in a $25B penalty order;
• predicted the results-day move would be large and would go against the earnings beat; Meta beat and fell 7%;
• cut its own conviction as the evidence came in, and stepped aside before the results.

We ran the same system on a consumer telehealth company and an offshore driller with a lot of debt. This post covers all three: what each swarm found, what it got wrong, and why we think swarms do this better.

heading: What we cover
Where the idea came from
What a swarm is
How much ground a swarm covers
Case study: Meta
Case study: Hims & Hers
Case study: Borr Drilling
The questions they'd ask management
Why a swarm finds what a single analyst misses
Do swarms get stuck?
What three swarms add up to

Where the idea came from

The inspiration is mathematics. Over the past year, OpenAI has shown what happens when you stop asking one model for one answer and let many reasoning agents work a hard problem together: propose, attack, check, repeat. That approach took AI from gold-medal-level performance at the 2025 International Mathematical Olympiad to contributing to open research problems.

• July 2025: AI models from OpenAI and Google DeepMind score at gold-medal level at the International Mathematical Olympiad. TechCrunch
• May 2026: OpenAI reports that one of its models disproved a central conjecture in discrete geometry. OpenAI

Mathematics has something investing doesn't. A proof is either right or wrong, and you can check it. Markets don't give you a proof checker. So we built the closest thing we could into every swarm:

• A built-in skeptic. One agent's only job is to challenge every assumption the thesis rests on.
• A scoreboard. The swarm makes dated predictions that can be proven wrong, and reality grades them.

What a swarm is

A Matterfact swarm is a research desk, not a chatbot.

• Human input. "Learn everything about the company, then build an investment thesis and questions for management." Runs until you stop it. Steer it at any time.
• Specialist agents. Dozens of agents, organised like a hedge-fund desk, divide the research and carry it out.
• Adversarial agents. A risk officer and outside points of view, such as a short seller, challenge every assumption in the thesis.
• Shared board. Task list (who is doing what), Messages (between agents), Feedback (people and agents), Decision log (every decision recorded).

columns: 2
One team, one board :: Agents divide the research and share every task, finding and open question on one board, the way an investment team shares one model. :: network
Every number has a source :: A number without a source is set aside until one is found. :: memory
Works around the clock :: It runs for weeks without being prompted, re-checking the thesis as new filings, prices and data arrive. :: clock
Dated, testable predictions :: Before an event, the swarm writes down what it expects, the date, and what would prove it wrong. :: target

How much ground a swarm covers

Each swarm ran for three to seven weeks and researched every part of the company it was given. What it builds is the full workspace an investment team would want on a name: the call, the valuation model, the evidence, what to watch, and how the view got there.

Meta: 7 weeks 1 Sep to 23 Oct; Hims & Hers: 7 weeks 3 Sep to 23 Oct; Borr Drilling: 3+ weeks 17 Sep to 9 Oct, non-stop.

What each swarm built for its company

• The call: Overview, IC memo, Our view vs the Street, Questions to ask
• Valuation: Valuation model (Summary, Workbook, Operating build, DCF & WACC, Equity bridge & dilution, Relative value, Scenarios & print, Sensitivity, Assumptions, Model checks), Peers & competition
• Evidence: Key debates, Bull / bear debate, Riskiest assumptions, Operating KPIs, Reported financials, Research library
• What to watch: Catalyst calendar, Company filings, Alternative data
• Behind the call: How the view evolved, Open questions, Track record & sources, Data room

Time spent matters less than where the signals came from. In all three case studies, the finding that mattered most came from a source most analysts never open: a lease footnote, a state court docket, a receivables facility filing, a mutual fund's bond valuations, a 35-year dayrate history pieced together from the annual reports of companies that no longer exist.

Case study: Meta

Meta Platforms (NASDAQ: META). A $1.5 trillion ad platform.

7 weeks of research, 1 Sep to 23 Oct; $704 → $573.77 fair value, 2 Sep to 23 Oct; Flat stance into and after results.

On 2 September, a day after it started, the swarm set out its first thesis:

• Long Meta.
• Probability-weighted fair value of $704 against a share price of $578.54, a 21.7% upside.
• Sized at about 4.5% of a model portfolio, with a market hedge.

Then the real work started.

What it looked at

• The physical AI build-out. It tracked 42 data-center commitments and checked them against utility planning filings, state utility-commission dockets and grid-connection queues. That told it when the capex turns into depreciation.
• 17 alternative-data series. Search interest for every Meta app and the Ray-Ban glasses, app-store rankings against TikTok, ad-buyer surveys, work-visa filings as a read on AI hiring, smart-glasses trackers, six currency pairs and small-business macro data.
• The legal tail, case by case. 42 legal and regulatory items were mapped by jurisdiction, probability and earnings impact.
• Who owns the stock. It built 39 views on ownership and what is priced in, including a solve for which scenario the share price was already pricing.
• Competitors. It read across Alphabet, Snap, Pinterest, The Trade Desk, Amazon and Nvidia, plus OpenAI's new ad business.

Meta: the swarm's fair value vs the share price ($ per share) Fair value fell as the swarm attacked its own assumptions, while the stock ran ahead. Share price is the hourly close.

META share price: $569.85 on 1 Sep 2026, $642.17 on 23 Oct 2026 (266 observations).

Marked events: 22 Oct 2026, Results: −7% next day.

It cut its own number four times in four days

The swarm then tested each assumption the thesis depended on. Each time it rebuilt that piece bottom-up, the answer came back worse than the thesis had assumed:

Meta fair value walk, 2 Sep to 23 Oct ($ per share) Four cuts in the first week, a full rebuild, then the results split into what got better and what got more expensive.

• Cost bridge (−$16.00). A 46.5% long-run margin needed costs other than depreciation to fall to 31.9% of revenue by 2030. Rebuilt line by line, they landed at 37.2%.
• Equity bridge (−$8.81). Reality Labs losses were under-deducted in three of the four scenarios.
• Exit multiple (−$19.99). A 15x exit multiple implied 3.22% perpetual growth, which only works if margins re-expand. 13x is consistent with the swarm's own margin path.
• Long-run margin (−$11.50). Three independent bottom-up checks put the long-run margin at about 41.5%, about 41.2%, and 38–41.5%. None supported 46.5%.

The signals that mattered, found later

1. The lease bill. The swarm built Meta's full infrastructure cost, including data-center leases that sit off the balance sheet until they start. On that full-cost basis, infrastructure goes from 9.8% to 24.8% of core-app revenue by 2030. The swarm used this to challenge its own margin assumption, and it fed into the cut to 43.1%. By the end of September, Meta had about $356B of signed leases that had not yet started. When the swarm rebuilt its model after the results, it made lease cost an explicit line, worth −$49.20 a share. Only capex and depreciation, at −$53.40, took more off. The swarm had found that bill seven weeks earlier.

2. The court case it refused to drop.

The swarm got the existence of the case right against its own doubts, and got the size badly wrong. It now carries the case at about $9.70 a share if upheld in full, pending appeal. On its read of the results call, neither management nor analysts asked about it.

3. The results-day call. Looking at the stock as a quant would, the swarm made a dated prediction. The results-day move would be bigger than ±6%, and it would not follow the direction of the earnings surprise. Meta beat on revenue by 2% and on earnings (excluding the legal charge) by 13%. The stock fell 6.98%.

4. The ad-volume ceiling. The swarm predicted that ad impressions growth would slow below 12% while users kept growing. Impressions rose 10%, and price per ad, +14%, did more of the work.

5. What kind of stock Meta is. A challenge the swarm set itself: "Meta is really an AI-capex trade." It tested this on market data and found Meta trades with ad stocks (90-day correlation 0.347) far more than with AI-capex stocks (0.112). The result held through the results (0.404 vs 0.114). That kept the hedge pointed at the right basket.

Conviction rose, then fell as the evidence came in

Meta swarm: conviction rose, then fell as each challenge landed (2 to 8 Sep) The swarm's conviction score at each stage of its first week. Higher is more confident in the long thesis.

Conviction peaked at 0.31 on 3 September and held near 0.25–0.30 for a few days. Then the challenges to the thesis kept landing, and each one took a step off. By 8 September, its own Monte Carlo gave the trade a 49% chance of winning, and the upside had shrunk from 21.7% to 4.5%.

On 30 September, with the stock at $738.79, a full rebuild put fair value at $621.60, and the swarm went flat. Three weeks later Meta beat and fell 7%. Today's fair value is $573.77. The stance is flat, with a level to buy near $499.

A research process that only ever adds conviction is a marketing department.

What else the swarm caught

What it got right

Meta swarm: how its 59 dated predictions have scored so far Of the 37 with a verdict, 18 were right (green), 7 partly right (amber) and 12 wrong (red). The other 22 never triggered, are still open or have not been scored yet (grey).

Of the 37 predictions with a verdict: 18 right, 7 partly right, 12 wrong.

The takeaway on Meta. The edge wasn't the headline number. It was the second-order bill behind it: the depreciation and leases that AI capex brings, found in lease footnotes and utility dockets weeks before they drove the revaluation. A specialist with a long memory kept a court case alive that turned into a $25B order. And conviction moved in steps as evidence came in, not with the share price.

Case study: Hims & Hers

Hims & Hers Health (NYSE: HIMS). A consumer telehealth company heading into its Q3 results.

7 weeks of research, 3 Sep to 23 Oct; $31.29 → $22.86 fair value, 3 to 27 Sep; 575 research notes written.

Same system, very different company: a consumer telehealth business heading into its Q3 results. In its first three and a half weeks, the swarm moved its fair value from $31.29 to $22.86.

Hims & Hers: the swarm's fair value walk, 3 to 27 Sep ($ per share)

Hims & Hers: the swarm's fair value vs the share price ($ per share) Fair value is shown at the dates the swarm re-based it; it has held since 27 September. Share price is the hourly close.

HIMS share price: $28.22 on 3 Sep 2026, $27.70 on 23 Oct 2026 (252 observations).

What it looked at

The swarm wrote 575 research notes, watched 1,000 SEC filings come through its feed, and tracked 33 alternative-data series:

• Demand: search interest for "hims", "hims weight loss", Wegovy and the GLP-1 category; app downloads; and web traffic to hims.com and forhers.com.
• Unit economics: revenue and gross profit per subscriber by month, and 24-month cohort curves for branded GLP-1 customers against the legacy hair and skin business.
• Pricing: the Hims weight-loss fee against its closest competitor's prepaid price, re-checked live on the websites.
• What the market is betting: short interest, the cost to borrow the stock, insider selling week by week, and the options-implied move for results day.
• Track record: revenue, EBITDA and gross margin against guidance and consensus for every quarter since late 2023.

1. It looked through the adjusted numbers

Hims reported $60.3M of adjusted EBITDA for Q2. The swarm split the $28.8M "acquisition and transaction" add-back into two parts:

• About $12M of one-time deal fees. Fair to add back.
• About $16.8M of pay for staff of acquired businesses, tied to them staying on. This comes from a $131.3M pool being expensed over years, so it's a recurring cost.

Then it treated stock-based pay, its payroll tax and an earn-out revaluation as real costs:

Hims & Hers Q2 FY2026: from reported adjusted EBITDA to cash EBITDA ($M)

Free cash flow was −$68.2M, a $128.5M gap to the headline. The swarm's summary: "the ramp is real on paper and thin in cash."

2. It corrected itself in the bullish direction too

The desk started with a common bearish view: Hims' gross margin was "really" 51%, flattered by non-cash add-backs. The 10-Q said otherwise:

• Reported gross margin was 63.8%, with no add-backs.
• $0 of stock-based pay sits in cost of revenue.

The 51% only appears if you move $95.5M of operations and support costs (12.7% of revenue) up into cost of goods.

Hims & Hers Q2 FY2026: where the “51% gross margin” came from (% of revenue)

It dropped the bearish argument, then pointed its skeptic at the right target: adjusted EBITDA.

3. It found mix shifts the thesis had missed

International was 12.9% of revenue in Q1 and 17.4% in Q2, heading toward 25%, not the ~5% one challenge assumed. That mix shift alone takes the margin on the non-weight-loss business from 76.6% to 73.7%.

The Australian acquisition (Eucalyptus) has no disclosed margin; the deal fell below the size that requires audited financials. The swarm estimated it four ways:

Eucalyptus gross margin: four ways to estimate a number nobody discloses (%) The swarm's adopted estimate in navy; the 52% the model had assumed in grey.

The estimates clustered around 44%, against the 52% the model had assumed. The swarm also caught the knock-on: lower Eucalyptus profit means the rest of the business earned more in Q2. So it raised the US margin rather than cutting twice. Net effect: −$1.22 a share.

4. It raised its value when the evidence required it

The swarm put a 63% chance on Q3 EBITDA landing at or below the guidance midpoint. Then it checked the company's record:

Hims & Hers adjusted EBITDA vs the midpoint of its own guidance (% beat) Never below the bottom of the range in 11 quarters; one small miss vs the midpoint. Zero is the guidance midpoint.

Hims had never reported EBITDA below the bottom of its range in 11 quarters. The swarm's bearish view turned out to be a pure gross-margin bet: on every other cost line, it was more optimistic than management. It blended three views:

Chance Q3 EBITDA lands at or below the guidance midpoint: three estimates, one blend (%) The estimate before the check in grey, the three views weighted equally in blue, the adopted blend in navy.

The chance of a weak quarter fell to 46%, and fair value rose $1.77. Within the hour, its skeptic challenged the upward move: the CFO said margin "dynamics" would persist, which means the trend, not a 64% level. That reading is worth about −$0.50, and the caveat was logged as a dated prediction.

5. It labelled its own work honestly

When the swarm moved to a stricter valuation model on 24 September, fair value fell about $6. It broke the drop down:

• About $710M of debt-like claims now came off equity value: deferred acquisition payments, an earn-out, a legal accrual and acquisition-linked pay.
• Stock-based pay was treated as a cash cost.
• Gross margin was rebuilt bottom-up at 61–62%.

Its verdict: about 60% of the cut was a stricter method, not a worse business. Most investors don't value Hims that way, so the swarm logged its own "overvalued" call as partly a matter of accounting convention.

The signals that mattered, found later

1. The bearish case was already priced in. The swarm read the options market's implied distribution for results day. Options priced a 36% chance of a 10%-plus drop; the swarm's own estimate was 35%. Being right about the downside would earn nothing, because everyone already expected it. That is the main reason the swarm has no trade on.

2. Hims beats on EBITDA but misses on gross margin. The swarm found consensus gross-margin estimates for 7 past quarters in a niche source. Hims missed the Street's gross margin in 5 of 7, by 1.9 points on average, while beating its EBITDA guidance midpoint 10 times out of 11. The pattern: the company delivers EBITDA by managing costs, not margin. That tells you which number to watch on 9 November.

3. Where the promised cash-flow turn could come from. First-half receivables jumped, and the swarm traced $347.4M of the build to rebates owed by drug makers, money that passes straight through. It then found a $400M facility, signed with JPMorgan on 1 July, that lets Hims sell receivables for cash. That could fund the second-half free cash flow turn management promised. So the swarm committed in advance to judge free cash flow excluding the facility.

4. Monthly billing raises cancellations. Moving branded weight-loss customers to monthly billing adds about 2 points a month to churn. That makes each customer less valuable.

5. The live price check. A month after first raising the question, the swarm re-read hims.com and five competitors' sites (22 captures). The price was still $39 for the first month, then $149. New: an anti-nausea prescription and AI support are bundled at the same fee, an early sign of how Hims will defend price.

What actually drives the value

Hims & Hers: what actually moves fair value ($ per share) Range of Base-case fair value as each input moves. The margin everyone debates (navy) is the smallest lever.

The debate everyone has about Hims, weight-loss gross margin, is the smallest lever on value. Acquisition cost and churn in the core business are four times bigger.

The takeaway on Hims. Precision isn't only finding errors. It's knowing three things: what the market has already priced (the downside), what actually drives value (acquisition cost and churn, not the headline margin), and what could flatter the next report (a receivables sale facility). Every outcome of the 9 November results is scored in advance. Whatever happens, the swarm can't talk itself into a story afterwards.

The full Hims story, including what the options market already expects and why the swarm has no trade, is in An Agent Swarm on Hims & Hers.

Case study: Borr Drilling

Borr Drilling (NYSE: BORR). An offshore driller with a lot of debt.

3+ weeks of non-stop research; $4.13 → $2.07 fair value, first build to audited model; Flat stance, held with high conviction.

Borr owns a fleet of offshore drilling rigs that work in shallow water. Its shares are a bet on two things at once: rig dayrates, and whether Borr can refinance its debt. This was the most intensive of the three: more than three weeks of non-stop research, 736 catalogued sources and 109 dated predictions.

What it looked at

• 35 years of dayrate history, adjusted for inflation and pieced together from the annual reports of companies that no longer exist.
• 373 real rig contracts, including original filings from 2001–2005.
• Cost benchmarks from 27 annual reports, to pin down what a rig really costs to run.
• Every rig, one by one: contract dates, customer, region, dayrate, and the idle gap before its next job.
• The full debt stack: the bond agreement, the revolver covenants, the maturity schedule, live bond prices, and mutual funds' filed valuations of the same bonds.
• Supply, insiders, peers and credit markets: new rigs leaving shipyards, insider-trading filings, peer dayrates, and high-yield spreads.

Borr Drilling: the swarm's fair value vs the share price ($ per share) The swarm found errors in its own model, rebuilt it rig by rig, then had it audited from scratch. Share price is the hourly close.

BORR share price: $4.38 on 17 Sep 2026, $4.78 on 9 Oct 2026 (119 observations).

Marked events: 25 Sep 2026, Flags its own model as wrong.

It found real errors in its own model, and said so

It fixed errors in the bullish direction too:

• It added back about $20M a year of reimbursable income the model had missed.
• It rejected a dilution charge that would have been counted twice.
• It cut maintenance capex from about $150–189M a year to the $60–70M the company guides to.

That last fix solved a puzzle. The equity model implied Borr's bonds should trade at 89 cents, but the market had them near 98. Instead of assuming the bond market was wrong, the swarm went looking for the bug in its own model.

The forensic signals it found later

1. Receivables that looked flat weren't. Reported receivables rose just 0.4%. Adding back $29.0M of receivables sold to a factoring company and an $18.2M allowance, they actually grew 25%. Days to collect were about 91, not the 73 reported. Factoring had flattered the half-year cash flow.

2. The maturity wall was ten times bigger than the summary said. The swarm's own summary had said "only $44.2M due before 2031". The debt note in the filing showed $502.3M of recourse debt due before 2032, more than Borr's $473.6M of liquidity.

3. Headline dayrates were inflated. Borr's stated 2026 average of $123k a day includes mobilisation fees and a gross-up on leased rigs. With those stripped out, the rate is $116–119k. Its South-East Asia rigs were earning about $97k, not the $115k the model assumed, a gap of roughly $28M a year of EBITDA.

4. The binding constraint isn't cash. In the Base case, the first test to fail is the revolver's 20% book-equity covenant (book equity was 26% at June), in late 2030. Liquidity runs out later. That changed which management question matters most: would the banks waive that test?

5. What shareholders get in a restructuring. The model assumed old shareholders keep $0.10 a share if Borr restructures. Looking at it as a regulator would, the swarm pulled actual offshore-driller restructuring plans from SEC filings. Old equity typically gets close to nothing directly, plus some warrants. The swarm re-marked the figure to about $0.03.

6. Bonds and stock disagreed. In early October the stock rallied to $4.61 while Borr's 2034 notes fell from 97.5 to 91.7 cents. Credit investors were moving toward the swarm's view even as equity investors moved away from it.

It audited itself, then rebuilt from scratch

In late September, the swarm reviewed its own work. It found 55% of a week's work had gone into re-arguing valuation method rather than finding new facts, and called some of it "process about process". So it rebuilt the model from the ground up:

• all 29 owned and 5 joint-venture rigs, quarter by quarter;
• every covenant tested on every path;
• a long-run dayrate spread across 19 outcomes, fitted to 35 years of history and 373 contracts;
• 414 assumptions, each with a source.

An audit of the rebuilt model, done from scratch, reproduced every result to six decimal places and passed all 116 checks. It also found four issues in the method, all of which were fixed. Fair value settled at $2.07.

Borr Drilling: chance of each long-run dayrate outcome vs its share of fair value (%) Half the probability sits where the equity is worth almost nothing; most of the value sits in the right tail.

In half the outcomes, debt eats almost all the value. Nearly two-thirds of fair value comes from the 17.5% chance that dayrates reach $150k or more. The shares are a long-dated option on the rig cycle.

It talked itself out of its own short idea

With fair value at $2.07 and the stock at $4.61, a short looks obvious. Then the swarm ran the same model on Transocean, Valaris, Noble and Seadrill. It valued every one at 0.3–0.7x its share price. So "worth less than the price" said something about the model, not about Borr. The swarm dropped the short label and now rates Borr "high-conviction flat", with dated events that would change that: a Pemex contract award, the October–November re-contracts, and the 19 November results.

It has logged 109 dated predictions, 13 of them scored so far, with an accuracy score of 0.137 (the Brier score, where lower is better). Always guessing 50-50 would score 0.25. For example, it put an 82% chance on the rig Norve not finding work within 30 days of its contract ending; its next job starts on 15 December.

The takeaway on Borr. The real advantage is catching your own mistakes at scale. Several of the errors were worth more than a dollar a share. The forensic finds came from documents most models never touch: factoring disclosures, the debt note, restructuring plans on SEC filings. And the swarm knew when not to trade: when its own model said every driller was overvalued, it treated that as a fact about the model.

The full Borr story, including the week it went in circles and what the audit found, is in An Agent Swarm on Borr Drilling.

The questions they'd ask management

Each desk writes its own questions for management and industry experts as it goes. Each one is built to settle a specific debate.

Every question on the sheet comes with four things:

columns: 2
What we believe now :: The desk's current view, with the number. :: evidence
Why it matters :: The dollars per share the answer could move. :: target
What answer would change our view :: Written down before anyone asks. :: lock
When to ask it :: A one-on-one, the results call, or the next conference. :: clock

The questions also get rewritten. As the research moves on, they get narrower, more numeric and harder to dodge.

Meta: from generic to unanswerable-without-a-number

The swarm's first draft, written on 3 September, read like most analysts' notebooks:

Meta, First draft, 3 September

"What incremental revenue/margin do you underwrite for the 2026-27 capex ($130-145B), over what payback period, and what ROIC hurdle governs the datacenter build?"

Management deflects that kind of question on every call. The swarm threw out the whole list and rebuilt it. Here are three of the questions that replaced it:

Meta, for management (About $44 a share)

"On the Q3 call last year you gave a directional view of the following year's capex before the formal guide. Will you do the same for 2027 this quarter, and can you walk us through which components drive the change: servers, data-center shells, network, and third-party capacity such as the Firmus agreement?"

This one points to something management has already done, so "we don't guide on that" doesn't work as an answer. It's tied to about $44 a share of value.

Meta, for management

"In July you said you expect 2026 operating income to be above 2025's. That was before the roughly $10 billion settlement accrual. Should we read the commitment as including or excluding that charge?"

This catches a promise the company made before a $10B charge landed. Including the charge, the promise means second-half operating income above about $51.6B.

Meta, for management

"Muse runs each user in a dedicated private virtual machine. How do you think about the cost to serve a Muse user compared with a Meta AI user, which cost lines carry it, and what would have to be true for the free tier to be margin-neutral?"

This goes after a cost that nobody on the Street was modelling.

Hims & Hers: questions built on the company's own words

Every Hims question starts from something management said, and asks it to pick between two readings:

Hims & Hers, for management (About $9M of EBITDA per margin point)

"On the Q2 call you said you expect the gross margin dynamics of the second quarter to persist through the second half. Should we read 'persist' as the Q2 level of about 64% being the right reference for the back half, or as the same mix forces continuing, which on our math takes the blended margin lower once Eucalyptus is in for a full quarter?"

This is the swarm's single biggest debate, boiled down to one word in one sentence from the CFO. Each point of Q3 gross margin is worth about $9M of EBITDA.

Hims & Hers, for management

"You hold new investments to a payback of a year or less. Do branded weight-loss customers, acquired at today's prices and on monthly billing, meet that standard on their own gross profit, and if they did not, what would you change first: price, marketing intensity, or the weight of weight loss in the funnel?"

This holds management to its own investment rule. On the swarm's numbers, a branded weight-loss customer returns only about 0.7x its acquisition cost over 24 months.

Hims & Hers, for management (±10% CAC moves about $13 a share)

"Search interest in the Hims brand is down about a fifth year on year on our measurement, yet you describe years of brand investment lifting acquisition in lower-cost channels. How do you reconcile the two?"

This sets the swarm's own alternative data against management's story. It goes after customer acquisition cost, the biggest driver of value: ±10% moves fair value by about $13 a share.

Borr Drilling: the question that sharpened itself three times

Borr's sheet shows the refining most clearly. Here is one question about contract-cancellation risk, as it changed over successive revisions.

Borr Drilling, for management, contract-cancellation risk

First version:

"Which of today's firm contracts carry termination-for-convenience or national-oil-company statutory termination rights, what early-termination fee applies, and how many suspension or termination notices has Borr received since 1 July 2026?"

After more digging, the swarm added:

"…Separately: who was the sanctioned customer behind the two 2025 terminations, and is any of that exposure linked to the Pemex conduit?"

After it found the answer itself, the question became a confirmation:

"…Separately, can you confirm the two 2025 sanctions-related terminations were the Hild and Odin contracts with Fieldwood Energy Mexico, and does any other current contract carry sanctions-termination language?"

The counterparty and rig names in the last version come from the swarm's own research. Borr has not confirmed them.

That's the pattern you want. The swarm doesn't spend management's time on anything it can find on its own. It uses the question to confirm what it found and to probe what's left.

The other Borr questions are just as specific:

Borr Drilling, for management

"On the same dayrate-equivalent basis you used for 2026 (73% covered at about $134k a day), what share of 2027 available rig-days is now contracted, and what is the average dayrate on that 2027 book excluding mobilisation fees?"

Borr Drilling, for management

"On South-East Asia, with Skald, Mist, Idun, Saga, Gunnlod and Thor coming off contract between November and April, how do you avoid bidding your own rigs against each other, and which of them have follow-on work in sight?"

The Borr sheet also prunes itself:

Borr Drilling: what happened to its 94 questions 30 were dropped as the model moved on and 13 were answered by the research before anyone needed to ask. The rest stay live.

A question list that never changes is a template. These lists keep getting rewritten as the research moves.

Why a swarm finds what a single analyst misses

1. Many specialists, one memory

One agent's finding becomes another agent's test:

• Borr's credit work on bond prices exposed a bug in the equity model's capex.
• Meta's legal-risk work kept the New Mexico case on its list after nearly dropping it twice, so the swarm was ready when the $25B order landed.
• Hims' alternative-data work found brand search down a fifth, which became a challenge to the Q4 earnings ramp.

2. A skeptic with nothing to defend

One agent's only job is to find the weak points in the thesis:

• It broke Meta's exit multiple.
• It attacked Hims' upward revision within the hour.
• It pushed Borr into the peer test that killed the short.

3. Triangulation instead of a single guess

• Eucalyptus' margin was estimated four ways.
• Meta's long-run margin was checked three ways.
• Borr's restructuring value was estimated three ways.

When the methods disagree, the swarm says so.

4. Commitments made before the event

224 dated predictions across the three swarms, each with what would prove it wrong. The swarm can't reinterpret a miss afterwards.

5. Breadth of source types, not just more of the same

Utility-commission dockets, state court dockets, lease footnotes, mutual-fund bond marks, options-implied distributions, restructuring plans and 35-year industry histories. The signal that moved each case came from one of these.

Do swarms get stuck?

Yes, sometimes. We've watched it happen, and the important part is that the swarms notice. We've seen three patterns.

1. Method loops

Borr spent most of a week re-arguing how to weight inputs that hadn't changed. Here is the range of fair values it had on the table:

Borr Drilling: the range of fair values on the table, before and after the rebuild ($ per share) Wide and wandering while the desk re-argued method; narrow once every change had to trace to a new fact.

From 25 to 29 September, the candidates ranged from $0.86 to $4.10. That's a desk circling, not converging. The swarm's review of its own work caught it, and the rebuild and audit broke the loop. After that, every change to fair value traced to a specific new fact: a contract award, a bond term, a restructuring precedent. The range narrowed to $2.07–2.65.

2. Data droughts

Over the Labor Day weekend, the Meta swarm had nothing new to read. It said so explicitly, stopped re-deriving its numbers, and held $647.70 rather than inventing motion.

3. Plateaus before a catalyst

Hims' fair value hasn't moved since 27 September. Most of the questions that would move it can only be answered by the 9 November results. We slowed it to a daily run.

Is the question list a treadmill?

On Meta, the swarm answered 133 questions and raised 131. Is that progress or running in place? The count alone can't tell you. What tells you is whether the answers change anything. We use three tests:

• Does a fair-value move trace to new evidence or to a change of method? The swarms label which one it is. Hims labelled 60% of its biggest cut as method.
• Did an answered question change an input? Meta's lease work changed the margin, then became its own line in the model.
• Did a prediction get graded? Reality is the one judge the swarm can't argue with.

How swarms get unstuck

• Reviews of their own work, with numbers rather than impressions.
• Challenges from outside points of view, such as a short seller or a regulator, that question the framing, not just the figures.
• Rebuilding when the foundation is wrong, rather than patching it.
• An audit of the rebuilt model, done from scratch.
• A human owner who can re-prioritise.

A swarm can't learn what isn't public yet. When it runs out of new facts, the right behaviour is to say so, hold, and have every possible answer pre-scored for when the facts arrive. That isn't being stuck. It's discipline.

What three swarms add up to

Together, these three swarms made 224 dated predictions, wrote 221 questions for management and industry experts, and catalogued more than 1,600 sources and research notes.

The result is a research desk that never sleeps, shows every step of its reasoning, admits when it's stuck, and gets more honest the longer it runs.

The Meta, Hims & Hers and Borr Drilling figures above are research outputs from Matterfact's agent swarms, shown to illustrate how the system works. They are not investment advice or a recommendation to buy or sell any security. Any trades or sizing described are hypothetical model outputs, not actual trades.