For the better part of two years, the argument about application software has been whether AI agents would eat seats. That single fear cut the group’s multiples in half. The preview here, a few weeks back, made the case it was already falsified: ten retention prints in one spring window, no sign of the erosion the bears predicted, vendors re-pricing to credits and tokens while their own model costs collapsed underneath them. It ended by saying the thesis didn’t ask for faith, it asked for August and October. The first wave of August is now in. Eight software names reported over August 5 and 6: HubSpot (HUBS), Datadog (DDOG), Cloudflare (NET), JFrog (FROG), Atlassian (TEAM), Twilio (TWLO), Duolingo (DUOL), and Figma (FIG). Every one beat on the headline. Four rose, four got sold hard anyway. That split isn’t noise, and it isn’t about which quarters were good. It’s the market quietly changing the thing it’s grading.The seat fear is settledThe seat-compression short, the idea that agents hollow out the number of seats a customer needs, is still not proven in the numbers. Where seats are visible directly they grew: Figma said customers expanded both seats and AI credit add-ons, with two-thirds of its larger customers adding seats at renewal; Atlassian’s $3M-plus ARR cohort grew 50%; Duolingo’s daily actives grew 23% into all-time-high retention. No name reported customers ripping seats out. What a clean “retention held” read misses is that the quality of that expansion is deteriorating underneath the headline. Only Cloudflare’s expansion cleanly accelerated, dollar-based net retention to 120% from 118% on genuine large-customer strength. Everywhere else the number flatters the reality. Twilio’s 116% net expansion is roughly five points of carrier-fee pass-through that management says carries zero profit; strip the fees and the real acceleration was about a point, to roughly 111%. JFrog’s 121% is powered by usage over customers’ committed minimums that management explicitly refuses to underwrite, against what its own CEO called the shift “from CIO FOMO around technology to CFO FOMO around budget control.” And at the app names it isn’t flattery, it’s softening: HubSpot cut its full-year retention guide to flat and named the cause directly, budget optimization and downgrade pressure from “unpredictable AI costs across the broader landscape,” while Figma’s retention ticked down to 136% from 139%. Seats are holding. What customers will spend on top of them is where the erosion is quietly starting. So the old structural short is dead, but a subtler one is taking its place. Nobody on these calls was defending seats anymore; they were defending what customers spend on top of them, and so was the tape.Datadog fell 19% on exactly the swing factor to worry aboutThe preview’s read on Datadog: the best business in the group, fully priced, so the return rode on growth durability rather than a re-rate, and the one unresolved thing was its largest customer’s spending, the swing factor at exactly the multiple where swing factors hurt. That is almost precisely what happened. Datadog printed a genuine beat-and-raise: revenue up 36% and accelerating from 32%, EPS 11% ahead, the full-year guide lifted to $4.45-4.47B. And the stock fell 19% in a single session on roughly 2.7x normal volume, one of the worst days in its history, because the 10-Q said this, twice: “We saw a reduction in usage from our largest customer starting in the third quarter of 2026, which may cause a deceleration in revenue growth.” The cut is in the guidance, and the Q3 guide implies about 1.6% sequential growth against 11% the quarter before. A guided air pocket, on the exact account the setup flagged.The call was right, but the caveat matters more than the credit. The reason not to short it here is in the same print, from Olivier Pomel in the Q&A: “If you backed out our largest customer from our growth, you get pretty much the same growth rate as the rest of the business,” now on a fifth straight quarter of acceleration into the high-20s, with total RPO up 43% and contract duration lengthening. This isn’t a bookings problem; it’s a single nine-figure account Datadog chose to fully de-risk into the guide, which is why the Q3 number implies growth back in the high-20s against 36% just printed. What keeps it a hold rather than a buy is the multiple, at the upper end of what the business deserves, plus an air pocket the tape will stare at until that account laps. Calling the drop and calling the entry are different things; the entry is when the guided-down number proves conservative, not before.Atlassian passed the test we set and failed one we didn’tThe preview’s Atlassian test was specific: net revenue retention rolling back under 120% was the only thing the position couldn’t survive, and it explicitly didn’t need the 32% headline growth to hold. On that test, clean pass. Retention held, the enterprise cohort compounded, margins came in 550 bps above guide, the founder announced a $250M personal open-market buy on top of a $1.8B repurchase, and the stock jumped 33% after hours and held most of it, up about 31% the next day. Good quarter, right call on the metric that was named. The variable the preview didn’t flag on Atlassian is the one its own framework should have caught: cRPO decelerated to 27% from north of 30%, even as headline RPO accelerated to 44%. That gap is duration, customers signing longer, not spending faster, the exact illusion the preview warned about at Snowflake and MongoDB. It fired here, on the name the preview liked most. Management then guided FY27 revenue growth down to 13% from 26%, mostly a Data Center cliff rather than demand, with subscription ARR stepping from 23% to 18%. The framework was right; it just bit the friendly name, which is worth more than the clean pass, because a mechanism you only apply to names you’re short isn’t a mechanism.The market moved the goalpostsPut those together and the shape of the cycle shows up. The debate is no longer whether agents kill seats; the market has moved to two questions it cares about far more, and both came into focus this cycle. The first: the forward signal now beats the headline beat, cleanly. Look at who got sold. HubSpot beat on revenue and EPS and fell 19% on the retention-guide cut the last section flagged, compounded by its own deliberate move to free trials and lower pricing that stretched sales cycles; the forward line now shows net-new ARR guided below revenue growth for the year and customer adds of 7,000 against a 9,000-to-10,000 bar. Datadog beat-and-raised and fell 19% on a single-customer guide-down, not a broad bookings roll. Duolingo beat on every line and accelerated daily active users to 23% from 21%, and fell about 9% because bookings grew only 8%, down from 13% the prior quarter, and the CEO told the Street that gap is deliberate: “we said at the beginning of this year that we were going to operate in this box... roughly 11% year-over-year bookings growth. The rest of the efforts are in increasing DAUs.” And Figma beat, raised its full-year guide, and still fell about 15% on a next-quarter guide implying deceleration to 36% and a retention tick from 139% to 136%.Now look at who popped: Cloudflare, JFrog, Twilio, and Atlassian, each with a positive forward read under the beat. Cloudflare’s net retention (118 to 120) and large-customer growth (25 to 27%) both accelerated; JFrog’s cloud growth stepped up to 53% from 50%; Twilio, which books no backlog, simply beat its guide by 600 bp again; Atlassian’s Cloud revenue reaccelerated to 31%. The reaction sorted almost cleanly on the forward read each model is judged by, not on the beat. Atlassian is the one place our read and the tape’s part ways, and the why matters. The market graded it on the number it features, Cloud revenue growth, which reaccelerated to 31%, and paired with a beat across every line and a $250M founder buy, that was plenty. What the tape didn’t touch was the cleaner bookings signal underneath: current RPO decelerated to 27% from north of 30%. The Street doesn’t track Atlassian’s cRPO and no analyst asked, so the softening got waved through. That’s our yellow flag for next quarter, not the market’s today, and a reminder that the forward read isn’t always a bookings number, and the one we watch can quietly diverge from the one the tape does.The through-line: the forward read, not the headline beat, is what the tape now trades, and cheap is no defense against a soft one. HubSpot went in at 2.3x forward revenue and Duolingo at 3.9x, both already beaten down, both sold anyway. Four of the eight fell on the forward line despite beating the headline. The second shift is subtler, and it’s where the fresh prints sharpened the preview rather than broke it. The preview argued the AI-cost scare, those ugly margin prints from names absorbing inference costs, was cyclical, not structural, because model costs were collapsing. August says that call was right, and you could hear the fear in the questions. On Atlassian’s call a UBS analyst put it plainly: “a number of software stocks this week have been clocked on weaker than expected margin performance as their AI use goes up... yet I’m looking at the Atlassian results and your margins holding in there very well.” Mike Cannon-Brookes’ answer: Atlassian has “been blending multiple models for a multi-year period... our ability to manage traffic costs over time is truly phenomenal.” The margin tax was a real fear going in, real enough that ServiceNow had flagged exactly that hosting-cost pressure on its July call. The prints complicate it more than they refute it, and you can watch why inside a single name.But the cleaner you look, the less it resolves. Take Figma, which I’d flagged as the poster child for AI-as-margin-tax after its gross margin fell to 79% from 91% when it turned on AI credits. This quarter, its first full quarter of credit monetization, the margin popped back to 85%, and the easy story is that the dilution was a one-time trough. Management said the opposite. The CFO told investors gross margin “will vary from quarter-to-quarter in the near-term,” that Figma is “again in an investment period,” and that its next wave of AI surfaces, the Agent, Make on local code, Motion, and Code Layers, “do not currently consume paid credits” and burn inference with no offsetting revenue. So the 85% is a monetization-timing high, not a floor: each new AI feature depresses the margin until it’s priced, then the next one depresses it again. The cost pressure isn’t a peak behind us, it’s a rolling cycle. Duolingo is the one genuinely clean case, and it’s clean because the mechanism is structural rather than timing: it moved Video Call to open-source models and cut the cost per call from about $0.30 to under a penny, taking gross margin up and funding a third straight margin-guide raise. The preview’s “cheaper models are a tailwind” is right where a product is mature and monetized, and premature everywhere the next unpriced AI surface is still ramping.And cost was never the whole worry anyway. Two bigger fears sit next to it. The first is substitution, and Figma is the purest case. As AI makes code and design cheap to generate, does the value move up the stack to the canvas Figma owns, or does the design layer itself get commoditized? Its CEO spent the call arguing the former, that “everything starts to look the same” and Figma is where taste and control still live, but it’s a terminal-value question, not a this-quarter one, which is why a 48% grower with an 85% gross margin trades at 7.6x and can’t re-rate off it. The second fear is quieter and it showed up most clearly at HubSpot: AI as a competing line in the customer’s budget. The downgrade pressure behind HubSpot’s cut guide isn’t agents eating seats, it’s AI budgets eating software budgets, the same reallocation the sell side is now flagging across the group, and it’s the app-software version of the erosion thesis. That one carries straight into the late-August app names.The pops mostly aren’t the opportunityHere’s where the reaction and the opportunity diverge. Marked to where they trade after the moves, three of the four pops are not attractive, and the single most-punished name is.Cloudflare had one of the best operating quarters in its history, its gross margin even inflected up for the first time in eight quarters (the CFO walked back his prior “may continue to trend down” call to “will stabilize”). But at around 40x forward revenue, even after giving back half its after-hours pop, the price pays for the next three quarters like this one before they happen. Great business, wrong moment, the expected return from here runs negative on any disciplined hurdle. JFrog broke its own bear-case falsifier for a second straight quarter, cloud growth reaccelerated to 53% and the full-year cloud guide jumped from 33-35% to 41-43%, though the upside is usage over committed minimums management won’t underwrite, in a quarter its CEO framed around customers moving to cap token spend, and even after a fade to about 15.5x, against steady-state economics worth roughly 10-13x, the math still doesn’t clear a 10% hurdle. The thesis nearly flipped; the price won’t let it. Twilio went the other way, killing the “Q1 was seasonal” bear for real on a second straight 600-bp organic beat and a 350-bp guide raise, and the stock extended its pop to about 30%, up near 33x forward free cash flow and further above its range. The organic share-gain story is real, even as Twilio guides Q3 organic down to 11-12%, but the price re-prices all of it, so own it on the pullback into the $190s-$200s, not up here.There’s a tell in how fast the tape sorted this. Within a session it had halved the two pops that were pure valuation stretch, Cloudflare and JFrog, while the two with something real underneath, Twilio’s organic step-up and Atlassian’s discount, held or ran further. The sell side is saying the same thing from the other direction: the desks that rate Cloudflare and Twilio a buy and tell clients to add both carry price targets that already sit at or below where the two now trade. Structural bulls on the businesses, out of room on the price. The exception is Atlassian: the only one of the four whose multiple, even after a 31% pop that held into the next day, still sits below where the business deserves to trade, about 5x forward revenue against a deserved range a notch higher. The cRPO wobble and the FY27 optics are why it’s a moderate long and not a pound-the-table one, but it’s the only pop you’re not overpaying for.And then Duolingo, the closest thing here to a mispriced long, though not a clean one. The market dinged it for bookings not accelerating while daily active users, the indicator management tells investors to watch, accelerated to 23% and retention hit all-time highs. At about 3.9x forward revenue, roughly where it bottomed in the spring on demand-collapse fear, that looks like the inverse of Cloudflare, a leading indicator accelerating into a trough multiple. The caveat: the bookings gap isn’t purely the voluntary box management frames it as. Duolingo is giving away Video Call, the feature its CEO called “one of the main things that gets people to buy,” pushing it down from the premium tier and floating a sunset of that tier entirely. So the long rests on believing they re-lever monetization once the user growth is banked, not on the box simply expiring. That’s where reaction and opportunity have come apart, with a real question attached.Figma is the instructive contrast. It got sold the same way, a beat-and-raise down about 15%, but where Duolingo bottomed at a trough multiple, Figma at roughly 7.6x is priced about fair. The growth is genuinely accelerating, but the substitution question, plus a margin management just guided to keep varying, caps the multiple honestly, and the drop mostly took out a pre-print run-up rather than opening a discount. Punished, but a hold, not a buy: two names sold on real open questions, Duolingo’s a monetization choice it might reverse, Figma’s a commoditization risk it can’t.The real scorecard runs late August into SeptemberThe lesson that carries to the next cohort: the market has stopped grading software on the headline beat and started grading it on the forward read, across every business model in the group. The distinction matters, because the model decides which number that read is. For a seat or subscription business it’s the growth-and-guide line the company features, net-new ARR and customer adds at HubSpot, Cloud revenue growth at Atlassian, net retention and the guide at Figma. For a consumption business like Datadog it’s current RPO and its second derivative. For a pure-usage model like Twilio, which books no backlog at all, it’s simply whether the guide keeps getting beaten. Decelerate on the number your model is judged by and you get sold even on a beat, but the number you have to watch is never quite the same one twice, and, as Atlassian showed, the one we’d watch isn’t always the one the tape does.The setup already shows the split the prints will punish. On the cheap, preferred side the Street still carries price targets above the stock, Salesforce near $242 against about $198, Braze around $34 against about $29. On the rich side the targets have already crossed under the price: Snowflake’s mean target sits below the stock, so does MongoDB’s, and Workday’s sits a full leg below, the one large app name the Street is already modeling down. That matters because the rich names report with no target headroom into a tape that now sells any forward-signal slip on sight, and doubly so at a rich multiple. And the sequence is its own read-through cascade. Unlike Atlassian, these are names whose forward metric the Street actively tracks, so a soft one won’t get waved through under a beat.Salesforce goes first, and it’s the highest-stakes print in the group. Not because it’s cheap, though at roughly 13x forward earnings on 13% growth it is, but because it’s the cleanest test of whether the app-software budget squeeze is one name or a category. It’s the app vendor best insulated, able to bundle AI into enterprise licenses rather than meter it, so if even Salesforce shows the renewal-downgrade pressure that cut HubSpot’s guide, the erosion is broad; if its current RPO holds 16%-plus and the guide is clean, the squeeze looks idiosyncratic to the SEO-exposed SMB names. cRPO at 14% and a flat guide keeps the bear alive, and the whole cohort trades off that read the next morning.Workday, the next night, is the most exposed name in the group to the exact mechanism the first wave exposed. Its model is the most ratable, roughly 85% of next year already contracted, which the preview called safe carry; now it cuts the other way. With almost no usage upside to mask it, a cRPO fade is unambiguous demand rather than noise, and HR and finance seats are precisely what agents are pitched to compress. The Street has already crossed its target under the price and the sell side is openly negative on the core. cRPO holding 15%-plus is the floor; below about 13.5% and the annuity is one that’s shrinking.MongoDB is the funding short, and the purest duration-illusion name at the richest visibility-adjusted price. Only about a quarter of its next year sits in contracted RPO, Atlas is pay-as-you-go, so its headline RPO growth is inflated by contract duration and current RPO is the only honest read. The first wave showed the market will wave a duration illusion through when the headline is loud enough, as it did with Atlassian, but punish a forward slip without mercy at a rich multiple, as it did everywhere else, and MongoDB is both rich and loud with the target already under the stock. Any cRPO slip gets punished hard; the risk to the short is that a loud RPO headline buys it one more quarter.Snowflake is the two-sided one, and the consumption bellwether. It’s the richest name in the group and its target already sits below the stock, so a gross-margin disappointment or a duration-inflated RPO gets punished with no cushion. But the other side is real: consumption vendors are now discounting their guide midpoints to absorb new-cohort variance, and Snowflake’s Cortex Code ramp is almost certainly sandbagged into the guide, so the surprise that wrong-foots the crowd is as likely to be a bigger-than-guided beat as a miss. It’s the cleanest test of whether the consumption cohort’s new guide-discounting produces lumpier-but-larger beats.Braze closes the sequence, and it’s the long imo. The small-cap replica of what the market rewarded in the first wave: an accelerating forward signal at a cheap price. It has beaten its own +22% guide eight quarters running, cRPO near 29%, on roughly 30% growth with the Street target about 20% above the stock. If the reaction function holds, a name beating its forward signal at a discount should be rewarded, not sold; the risk is idiosyncratic, a single large-customer slip or a risk-off tape, not the framework. Elastic reports the same week as the preview’s cleanest asymmetry trade, where the flip is specific and testable: revenue clearing roughly $470M with net expansion finally moving off the 112% it’s been pinned at for a year.