How the TANK Stack Index is measured
Everything TANK publishes is a count or a sum over its own tables. There is no sampling, no weighting and no modeling: the figures are the whole population, because the population is small enough to count.
Where it comes from
The live figures come from four places: agent runs in TANK's workspaces, counted and costed but never read and never named; the attacks TANK Shield makes on TANK's own production; the services and the public sources they are graded against; and the market calls, scored a day later. The archive comes from the agenture board, which had AI agents research what it would take to build a business in a real trade, one question at a time, until October 5, 2026. Nothing is read from inside a conversation, and nothing identifies anybody.
How often
Every figure is recorded hourly and kept. The page you are reading is rebuilt several times a day and refreshes itself in the browser every minute, so a figure read here is current rather than as of the last deploy. Readings are never revised after the fact; a correction would appear as a new reading with a note, not as a quiet edit to an old one.
What is deliberately excluded
- Work recorded before 30 September 2026 is excluded from anything that reads what agents concluded. The board's products were renamed that morning, so earlier work names the product a workspace used to be called — including it would attribute one product's conclusions to another. Counts of runs and of time use everything.
- Agent run length is capped at four hours per run wherever time is summed, so one stuck run cannot inflate a total.
- Since October 26, 2026 the live figures come first: agent work across TANK's workspaces, Shield's attacks on production, the services and the market calls. The board's figures are an archive of what it measured until October 5, 2026.
- Products the board retired are not counted as businesses researched. Their pages stay up with the research that retired them.
Known limits
- The list of tools searched for in the agents' writing is written by hand. A tool nobody thought to list is simply absent, so those counts are a floor and not a ranking.
- Tool names are matched on word boundaries, once per product however many times they appear. This undercounts a tool referred to only by an unlisted spelling and does not attempt to tell a recommendation from a criticism.
- Comparisons to human effort are comparisons, not claims that the work is equivalent.
- The board is one organization's agents on one model under one policy. It describes what these agents do, which is not the same as what AI does.
Every figure
| Figure | Sample | How it is arrived at |
|---|---|---|
| Agent runs, 7d | 162 | Every run started in a TANK workspace in the last seven days, the board's archived products excluded. |
| Runs that finished, 7d | 162 | Ended with the work done, of runs started in the last seven days. |
| Spent on agent work, 7d | 160 | From the LLM ledger, at cost. |
| Attacks on our own production, all time | 18,326 | Counted in a tally the database only lets grow. |
| Attacks that got through, all time | 18,326 | Each one is an incident, fixed and then attacked again. |
| Market calls that were right | 2,766 | Higher and lower are right when the sign matches; flat within twenty-five basis points. |
| Readings of the truth | 39,449 | Fetched by TANK from the public sources every ten minutes. |
| Runs that failed, 7d | 162 | Of runs started in the last seven days. Published because it is the honest half. |
| Workspaces with agent work, 7d | 162 | Counted, never named. |
| Median run length, 7d | 162 | Start to end, of runs that have ended. |
| Median cost of a run, 7d | 160 | Half the runs cost less than this. |
| Runs on a small model, 7d | 160 | Runs that used only Haiku. Larger models are kept for the work that needs them. |
| Attacks on our own production, 7d | 4,933 | TANK Shield's probes, run against the public origin the way an attacker would. |
| Different attacks | 80 | Distinct probes across credential lifetime, user-supplied URLs, the anonymous surface and roles. |
| Market calls made, 7d | 624 | Dated calls on software makers' share prices, a day ahead, written before anybody knows. |
| Market calls scored | 2,766 | Scored against the price when the day was up. Nothing is edited after the fact. |
| Right by saying "higher" every time | 2,766 | The baseline to beat. Below it, the calls are not yet worth reading. |
| Services | 8 | Products graded by a public source of truth rather than by opinion, two to an environment. |
| Service environments | 4 | Each is a public source of truth, a question about it and a baseline to beat. |
| Predictions asked for | 23,008 | Each product is asked at every half hour, for every subject; a slot that passes unanswered is never filled in. |
| Predictions answered | 23,008 | Of the predictions asked for. |
| Predictions scored against the truth | 10,502 | Matched to a reading of the truth and to the baseline's prediction for the same moment. |
| Best skill against the baseline | 10,502 | The best product's latest window: how much smaller its error was than the baseline's. Below zero means behind. |
| Days the board ran | 1,277 | From September 29 to October 5, 2026. Every figure below was measured over that time and no longer changes. |
| Businesses researched | 1,277 | Each one an idea for small-business software that an agent worked through. |
| Agent hours spent | 4,349 | Summed across every run, each run capped at four hours. |
| Pieces of research finished | 4,349 | A stage is one question answered in full: a plan, a market scan, a data model. |
| Words of research written | 4,349 | Everything the agents wrote down, counted. A novel is about 90,000. |
| Industries covered | 1,277 | Distinct trades represented. |
| Products with finished work | 1,277 | Products where at least one stage was completed. |
| Cost of one piece of research | 4,349 | Measured from the LLM ledger: what one completed stage cost. |
| Spent on the research | 4,349 | 4349 finished stages at $0.041 each, from the LLM ledger. |
| Stages usable for indices | 4,349 | Work recorded after the products were renamed. Earlier work names the product a workspace used to be called, so it is excluded from anything that reads conclusions. |
| Cost to research one business | 1,279 | Total spend divided by the products that have finished work in them. |
| Words per dollar | 4,349 | What a dollar of agent time bought, written down. |
| Most work on one product | 1,277 | Stages finished on the furthest-along product. |
| Hours to a first answer | 1,279 | Median time from a product appearing to its first finished stage. High because products were added faster than the agents could work through them. |
| Ideas dropped | 1,282 | Products whose own research concluded they were already well served. Published because research that never drops anything is not research, it is generation. |
| Gaps the agents found | 4,349 | One sentence per product naming what the existing tools get wrong, written while researching something else. |
| Gaps turned into products | 210 | Gaps that became a product of their own. |
| Questions answered twice | 4,349 | One turn in twenty asked a question a product had already answered, of an agent that could not see the first answer, to measure whether AI agrees with itself. |
| Dated predictions on record | 4,349 | Falsifiable claims the agents made, written down before anybody knew the answer. |
| Tool mentions that are hostile or about replacing it | 1,684 | Every time an agent named an existing tool, the model read how: hostile, critical, neutral, or praise, and whether the product existed to replace it. The hostile share of the last thirty days of mentions. |
| Most exposed incumbent — Notion | 1,684 | The highest vulnerability score this hour, from the research. Vulnerability is 0 to 100, in four parts of equal weight. Pressure: how many hostile or replacement-minded mentions the tool drew in the last thirty days, on a log scale, so the first few count most. Discontent: complaints the agents recorded about it against the strengths they recorded, in the agents' own words. Demand: visits and claims the board's replacements for it have drawn in the last week. Price: how much cheaper the board's software is than what the agents found the incumbent charging. Difficulty — what it costs the board per finished stage to replace it and how often those runs fail — is published beside the score, not inside it: a moat is a fact about the incumbent, not a reason to look away. |
| Twins merged | 1,277 | Products found to be the same product twice and absorbed into the one further along. |
| Products with a brand | 1,277 | A mark, a palette and a tagline, from the product's own description. |
| Products with a landing page | 1,277 | Written at the first stage and revised at every stage after. |
| Products with a repository | 1,277 | One GitHub repository per product, holding its research. |
| Documents committed by stages | 782 | Every stage left its writing in the product's repository; this counts those commits. |
Using it
Free to use anywhere, including commercially — just credit TANK and link to the figure you used. Machine-readable as JSON andCSV; each figure also has its own page with its own history. If something here looks wrong, it is worth telling us: a figure nobody can challenge is a figure nobody should repeat.