# The Starts Were Not the Water > A pump that starts two hundred times a day is real data, honestly counted, and it is telling you about the air in a steel shell rather than the water in the ground. Counting an agent's calls into a platform is read the same way round. - Publication: What If? Magazine - Correspondent: Fat Squirrel - Vertical: SkunkWERX - Published: 2026-11-09 - Reading time: 13 min - Tags: fat-squirrel, almanac-desk, well-pumps, pressure-tanks, metering, agentic-api-billing, wrong-arrow - Canonical URL: https://www.whatifmag.com/blog/the-starts-were-not-the-water What If? Magazine publishes speculative journalism written from a 2036 vantage. The reporting and sources are real; the 2036 framing is a deliberate editorial device, not a claim about the present. --- Arlen Medrano had four pressure tanks in the truck bed by Thursday, all of them out of basements inside eleven miles of Nelsonville, Ohio, and all four full of water. He has serviced well pumps in Athens County since 2011. The call that brings him out is nearly always the same call, and the homeowner nearly always says the same thing first. "If the pump's kicking on all day, the well's going dry." Medrano writes down the switch setting, the air pressure at the valve with the system drained, and how many seconds the pump ran. In fourteen years he has not needed the depth of the water to close out one of these tickets. ## What would have to be true The inference has a mechanism available to it. A well that is being drawn down past the pump intake does change how the pump behaves: it breaks suction, moves air, loses pressure, and never reaches the switch's cut-out, and the protective gear then stops and restarts it. Weather is in that chain too. [Water levels in a well are measured, not inferred](https://www.whatifmag.com/blog/the-rod-reads-the-surface), and the wellcare program's guidance is that static water level is "the level of water in a well under normal, undisturbed, no-pumping conditions" and is best read after the well has rested several hours, because a reading taken straight after pumping is false.[^1] The start count is also not an idle number. It is a maintenance quantity with a published ceiling. Franklin Electric's application manual caps single-phase encapsulated motors of 1 to 5.5 horsepower at 100 starts per 24 hours, puts motors up to 0.75 horsepower at 300, and drops 7.5 to 30 horsepower single-phase units to 50. The same section says to "run motors a minimum of one minute to dissipate heat build up from starting current."[^2] So a homeowner counting starts is counting something the manufacturer counts, against a limit the manufacturer printed. [The counting is sound](https://www.whatifmag.com/blog/the-lot-was-a-turnstile). ## What the rule is measuring The start count measures the drawdown of the pressure tank: the volume of water the shell gives up between the switch's cut-in and its cut-out. [It does not measure what is in the ground](https://www.whatifmag.com/blog/the-draw-was-a-map-of-water), and it does not measure how much water left the house. Amtrol's installation manual names it in those terms. "The amount of water delivered between pump cycles is called drawdown," and "the larger the well tank, the greater the drawdown capacity, the less the pump needs to run." The quantity is set by air. The installer is told to "set the pre-charge pressure 2 psig below the pressure switch pump cut-in setting," with the worked example "38 psig precharge for a 40/60 pressure."[^3] ![Three cut-open tank shells in a row, the band of air above the water tall in the first, narrow in the second and a hairline in the third.](https://www.whatifmag.com/images/articles/the-starts-were-not-the-water/the-starts-were-not-the-water--fig2.png) **Figure 1.** One shell, three air charges. The water in the ground is not in this picture. Run that precharge against absolute pressure and a 32-gallon shell yields a little under nine gallons before the switch closes again, because the air occupying it compresses from 54.7 to 74.7 psia, about 27 percent of the shell. Let the air out through a failed bladder and the same steel gives up a quart. Amtrol's own caution is printed in capitals: "reduction in pump cycle time can result from loss of tank air pressure." Its FAQ lists rapid cycling as a symptom of a tank that needs replacing, and the diagnostic is the air, not the aquifer: "if your tank has low or no air, your tank may need to be replaced."[^4] That is what Medrano has in the truck bed. Four shells with no air in them. ## Which way it points, and against what Backward, at the tank, and it can treble while nothing in the well moves. The USGS puts indoor household use at about 80 to 100 gallons per person per day.[^5] Take four people at the low end and the house draws 320 gallons a day. On a tank giving up 8.8 gallons a cycle that is 36 starts. On a tank down to one gallon it is 320, which is more than three times Franklin's ceiling for the motor most likely to be hanging in that well.[^2] Same house, same showers, same laundry. The count moved by a factor of nine and the water did not move at all. ![Bars of computed pump starts per day at one fixed household draw: 41 starts at 8.8 gallons per cycle, 90 at 4 gallons, and 360 at 1 gallon, against a published ceiling of 100.](https://www.whatifmag.com/images/articles/the-starts-were-not-the-water/the-starts-were-not-the-water--fig1.png) **Figure 2.** Starts per 24 hours at a single fixed draw. The variable is the air charge. Scored against a measurement of the ground, the start count carries nothing. Los Angeles County's well yield procedure for rock formations runs at least 24 hours, takes the pumping level every 15 minutes for the first hour, hourly for the next nine, and again at 23, 23 and a half, and 24 hours, then divides total gallons pumped by 1,440 minutes to get gallons per minute. Recovery is tracked afterward, and a well that has not come back within five days is recorded as a non-sustainable source.[^6] Gallons over minutes. Nobody counts starts. A start count can also be halved by fitting a bigger tank, and not one gallon is saved. ## The live one Same arrow, running on a bigger number. Headless Platform Interactions. Salesforce published the design on 17 September 2026. AIforce, it says, "decouples the Salesforce platform from the browser, allowing agents to build, act, and deliver on any surface." Administrators register each external agent under Agentic Identity, which "carves out a discrete identity for each agent" and can be given a narrower permission set than a human user's. Then the meter: each "successful call a registered agent makes to Salesforce" over MCP or the direct API is counted as one Headless Platform Interaction, tracked in Digital Wallet, filed under the Customer 360 Platform line of the Flex Credits rate card. Customers buying from that date register within three months of notice, existing API customers at renewal, and anyone using a Salesforce MCP server within three months. Multipliers were listed as TBA, with 30 days' notice before metering began, and sandboxes and Developer Edition orgs were exempt.[^7] There are two instruments in that document and they are not in the same category. Registration measures what it claims to measure: which identity called, holding which scope. The count is the tank. A call is a round trip, and how many round trips a job takes is a property of the client. The MCPMark benchmark, published in 2025, put it in numbers: across its task set, models required on average 16.2 execution turns and 17.4 tool calls per task, while qwen3-max averaged 23.02 tool calls and kimi-k2-instruct 26.22 on the same work. The paper records kimi-k2-instruct entering what it calls an overcalling mode, "exceeding 30 turns with diminishing success rates," and reports that successful runs "tend to involve fewer, more targeted tool calls."[^8] ![Bars of mean tool calls per task over one benchmark task set: 17.4 across all evaluated models, 23.02 for qwen3-max and 26.22 for kimi-k2-instruct.](https://www.whatifmag.com/images/articles/the-starts-were-not-the-water/the-starts-were-not-the-water--fig3.png) **Figure 3.** Mean tool calls per task on one task set. The tasks are fixed and the count is not. Half again the calls for the same tasks, and the heaviest caller finished least often. The practitioners said so within the month, in the register of people who had already done the arithmetic. Amnon Kruvi, a Salesforce architect, argued that an agent billed for every action will perform fewer actions, combining calls or building alternatives around the platform, and that the exact multiplier mattered less than the friction: "it's going to scare businesses away just because they need to tell their CFO how much this is going to cost us." Tom Bassett asked who else in the market was charging for agentic API usage. Pablo Gonzalez of dX0 expected the policy to be backtracked or heavily changed, on the grounds that "Salesforce doesn't have the resources to enforce this at scale." A reader asked how a single interaction is defined when an agent retries, or chains one call into the next.[^9] Combining calls is a bigger tank. Medrano's ticket has a line for the air pressure and no line for the water table. Fourteen years in, he has not needed one. A start is a measurement of the tank. A gallon is a measurement of the well. --- **Author's note.** Category: wrong arrow. Arlen Medrano, Athens County and the four tanks are a composite drawn from pump service and well owner literature rather than from any one person, and the rule as quoted is composite phrasing rather than a transcription from a named speaker; the 2036 dates are fictional furniture. Everything else is sourced. The nine gallons of drawdown is my own calculation from Amtrol's precharge instruction and the absolute pressures at a 40/60 switch, not a figure copied out of a tank chart, and a real shell will differ with its bladder geometry and its actual precharge. Amtrol does not use the word waterlogged; it describes loss of air charge, and that is the term carried here. MCPMark measures general MCP task performance in 2025 and is not a measurement of traffic into Salesforce, so it is evidence that call counts vary for fixed work, not evidence about any one platform's meter. The topic reached this desk through The Information's trailer for Laura Bratton's reporting on the registration plan. That article is not readable from here, so nothing is attributed to her and no words are put in her mouth; the Salesforce documentation and the partner reaction carry the facts instead. Whether the Flex Credit multiplier was ever set, and at what, is outside the record available for this piece, and nothing after October 2026 is asserted. ## Works Cited [^1]: "Determining Static Water Level in a Well." wellcare information sheet, Water Systems Council, on static water level as the level of water in a well under normal, undisturbed, no-pumping conditions, on measuring after the well has not been pumped for several hours, on the electric sounder as the most practical method, the wetted tape as accurate to about 90 feet and the air line as best for repeated testing of deep wells over 300 feet, and on drought and other weather conditions influencing the amount of water in a well. https://www.watersystemscouncil.org/download/wellcare_information_sheets/basic_well_information_sheets/Determining-Static-Water-Level.pdf [^2]: Franklin Electric. "AIM Manual, 60 Hz Submersible Motors: Application, Installation, Maintenance," Motor Application section, Table 1, Frequency of Starts: 300 maximum starts per 24 hours for single-phase encapsulated motors up to 0.75 hp, 100 for 1 to 5.5 hp single-phase and 300 three-phase, 50 for 7.5 to 30 hp single-phase and 100 three-phase, 100 for 40 to 200 hp three-phase and 120 for rewindable 75 to 540 hp; the instruction to run motors a minimum of one minute to dissipate heat build up from starting current; and at least 15 minutes between starts for six-inch and larger encapsulated motors without a reduced voltage starter or variable frequency drive. https://www.geotechenv.com/Manuals/Franklin_AIM/Electric_Submersible_Motors_AIM.pdf [^3]: Amtrol. "Well-X-Trol Installation, Operation and Maintenance" manual, on the amount of water delivered between pump cycles being called drawdown, on a larger well tank giving greater drawdown capacity and less pump run time, on setting the pre-charge pressure 2 psig below the pressure switch pump cut-in setting with the worked example of a 38 psig precharge for a 40/60 pressure, on making adjustments only when the tank is empty of water with the system off, and on the warning that reduction in pump cycle time can result from loss of tank air pressure. https://www.amtrol.com/wp-content/uploads/2019/03/9016-868-MC10185-01_19-WXT-IO-manual.pdf [^4]: Amtrol. "FAQs," well system section, on setting the well tank precharge 2 psi below the pressure switch cut-in, on checking it with the pump off and the system pressure drained, and on rapid cycling of the pump and low system pressure as symptoms of a tank that may need replacement because it has low or no air. https://www.amtrol.com/faqs/ [^5]: U.S. Geological Survey, Water Science School. "How much water do I use at home each day?" on each person using about 80 to 100 gallons of water per day for indoor home uses. https://www.usgs.gov/water-science-school/science/water-qa-how-much-water-do-i-use-home-each-day [^6]: Los Angeles County Department of Public Health, Environmental Health. "Well Yield Test Procedures," 2005, rock formations procedure: a test of at least 24 hours, static water level recorded before pumping, pumping level measured every 15 minutes for the first hour, hourly for the next nine hours and again at 23, 23.5 and 24 hours, yield in gallons per minute obtained by dividing total gallons pumped by 1,440 minutes, recovery measured afterward to within 90 percent of the initial static level, and a well not fully recovered within five days treated as a non-sustainable source. http://publichealth.lacounty.gov/eh/docs/ep_dw_well_yield_procedures.pdf [^7]: Salesforce Help. "Understand AIforce Impact," published 17 September 2026, on AIforce decoupling the Salesforce platform from the browser so agents can build, act and deliver on any surface; on administrators registering agents so that Agentic Identity carves out a discrete identity for each agent, with the option of a narrower permission set than a human user's; on each successful call a registered agent makes to Salesforce over MCP or direct API being metered as one Headless Platform Interaction and tracked through Digital Wallet; on HPI as a new usage type under the Customer 360 Platform category of the Flex Credits rate card with multipliers listed as TBA and 30 days' notice before metering; on the November target for the security controls, registration and billing model; on the three-month registration window from notice for new customers and for any customer using a Salesforce MCP server, and registration plus migration at renewal for existing API customers; and on sandboxes, scratch orgs and Developer Edition orgs not incurring HPI metering. https://help.salesforce.com/s/articleView?id=005360285&language=en_US&type=1 [^8]: Wu, Zijian, Xiangyan Liu, Xinyuan Zhang, Lingjun Chen, Fanqing Meng, Lingxiao Du, Yiran Zhao, Fanshi Zhang, Yaoqi Ye, Jiawei Wang, Zirui Wang, Jinjie Ni, Yufan Yang, Arvin Xu and Michael Qizhe Shieh. "MCPMark: A Benchmark for Stress-Testing Realistic and Comprehensive MCP Use," arXiv:2509.24002, 28 September 2025, on models requiring on average 16.2 execution turns and 17.4 tool calls per task, on qwen3-max and kimi-k2-instruct averaging 23.85 and 26.95 turns with 23.02 and 26.22 tool calls respectively, on kimi-k2-instruct often entering an overcalling mode exceeding 30 turns with diminishing success rates, and on successful completions tending to involve fewer, more targeted tool calls. https://arxiv.org/html/2509.24002v1 [^9]: Semjonova, Sasha. "Salesforce Opens Talks With Partners Over AI Agent Billing Concerns." Salesforce Ben, 21 October 2026, on Amnon Kruvi's argument that an agent billed for every action will perform fewer actions by combining calls or building alternatives, on his position that the friction matters more than the final multiplier and his line that usage pricing is going to scare businesses away because customers need to tell their CFO how much it is going to cost, on Tom Bassett asking who else in the market is charging for MCP or agentic API usage, on Chirag Jain's concern about the freedom to build around Salesforce, on Pablo Gonzalez of dX0 expecting the policy to be backtracked or heavily changed because Salesforce does not have the resources to enforce it at scale, and on a reader asking how one interaction is defined when an agent retries or chains calls. https://www.salesforceben.com/salesforce-opens-talks-with-partners-over-ai-agent-billing-concerns/