The TANK Stack Index
Edition 2026-11 · measured hourly · how
What TANK measures every hour: the agent work in its workspaces, the attacks it makes on its own production, the services graded against public sources, and the market calls scored a day later. After them, the research AI agents did on twelve hundred small-business ideas, kept as an archive. Every figure says what it rests on. Take any of it: CSV, JSON. Free to use anywhere, including commercially — just credit TANK.
- 162Agent runs, 7dEvery run started in a TANK workspace in the last seven days, the board's archived products excluded.
- 69%Runs that finished, 7dEnded with the work done, of runs started in the last seven days.
- $348.59Spent on agent work, 7dFrom the LLM ledger, at cost.
- 18,326Attacks on our own production, all timeCounted in a tally the database only lets grow.
- 1Attacks that got through, all timeEach one is an incident, fixed and then attacked again.
- 39%Market calls that were rightHigher and lower are right when the sign matches; flat within twenty-five basis points.
- 39,449Readings of the truthFetched by TANK from the public sources every ten minutes.
Live figures are recomputed every hour.
Agent work in TANK
Runs in every TANK workspace over the last seven days, counted and never named.
| Figure | Value | Sample | What it measures |
|---|---|---|---|
| Runs that failed, 7d | 22% | 162 | Of runs started in the last seven days. Published because it is the honest half. |
| Workspaces with agent work, 7d | 18 | 162 | Counted, never named. |
| Median run length, 7d | 6.9 min | 162 | Start to end, of runs that have ended. |
| Median cost of a run, 7d | $0.745 | 160 | Half the runs cost less than this. |
| Runs on a small model, 7d | 28% | 160 | Runs that used only Haiku. Larger models are kept for the work that needs them. |
Attacks on our own production
TANK Shield attacks the public origin the way an attacker would, every hour, and keeps the count.
| Figure | Value | Sample | What it measures |
|---|---|---|---|
| Attacks on our own production, 7d | 4,933 | 4,933 | TANK Shield's probes, run against the public origin the way an attacker would. |
| Different attacks | 80 | 80 | Distinct probes across credential lifetime, user-supplied URLs, the anonymous surface and roles. |
Services, graded by the world
Products whose answers are checked against a public source of truth and a baseline to beat.
| Figure | Value | Sample | What it measures |
|---|---|---|---|
| Services | 8 | 8 | Products graded by a public source of truth rather than by opinion, two to an environment. |
| Service environments | 4 | 4 | Each is a public source of truth, a question about it and a baseline to beat. |
| Predictions asked for | 23,008 | 23,008 | Each product is asked at every half hour, for every subject; a slot that passes unanswered is never filled in. |
| Predictions answered | 49% | 23,008 | Of the predictions asked for. |
| Predictions scored against the truth | 10,502 | 10,502 | Matched to a reading of the truth and to the baseline's prediction for the same moment. |
| Best skill against the baseline | 100% | 10,502 | The best product's latest window: how much smaller its error was than the baseline's. Below zero means behind. |
Market calls
Dated calls on software makers' share prices, a day ahead, scored when the day is up.
| Figure | Value | Sample | What it measures |
|---|---|---|---|
| Market calls made, 7d | 624 | 624 | Dated calls on software makers' share prices, a day ahead, written before anybody knows. |
| Market calls scored | 2,766 | 2,766 | Scored against the price when the day was up. Nothing is edited after the fact. |
| Right by saying "higher" every time | 53% | 2,766 | The baseline to beat. Below it, the calls are not yet worth reading. |
The research archive
The agenture board had AI agents research small-business software, one idea at a time. It was stopped on October 5, 2026. From September 29 to October 5, 2026. Every figure below was measured over that time and no longer changes.
| Figure | Value | Sample | What it measured |
|---|---|---|---|
| Businesses researched | 1,277 | 1,277 | Each one an idea for small-business software that an agent worked through. |
| Agent hours spent | 226.2 | 4,349 | Summed across every run, each run capped at four hours. |
| Pieces of research finished | 4,349 | 4,349 | A stage is one question answered in full: a plan, a market scan, a data model. |
| Words of research written | 715,505 | 4,349 | Everything the agents wrote down, counted. A novel is about 90,000. |
| Industries covered | 86 | 1,277 | Distinct trades represented. |
| Products with finished work | 1,279 | 1,277 | Products where at least one stage was completed. |
| Cost of one piece of research | $0.041 | 4,349 | Measured from the LLM ledger: what one completed stage cost. |
| Spent on the research | $178.31 | 4,349 | 4349 finished stages at $0.041 each, from the LLM ledger. |
| Stages usable for indices | 3,758 | 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 | $0.139 | 1,279 | Total spend divided by the products that have finished work in them. |
| Words per dollar | 4,013 | 4,349 | What a dollar of agent time bought, written down. |
| Most work on one product | 6 | 1,277 | Stages finished on the furthest-along product. |
| Hours to a first answer | 19.6 | 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 | 5 | 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 | 210 | 4,349 | One sentence per product naming what the existing tools get wrong, written while researching something else. |
| Gaps turned into products | 5 | 210 | Gaps that became a product of their own. |
| Questions answered twice | 70 | 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 | 122 | 4,349 | Falsifiable claims the agents made, written down before anybody knew the answer. |
| Tool mentions that are hostile or about replacing it | 0% | 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 | 28 | 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 | 5 | 1,277 | Products found to be the same product twice and absorbed into the one further along. |
| Products with a brand | 1,277 | 1,277 | A mark, a palette and a tagline, from the product's own description. |
| Products with a landing page | 1,277 | 1,277 | Written at the first stage and revised at every stage after. |
| Products with a repository | 1,277 | 1,277 | One GitHub repository per product, holding its research. |
| Documents committed by stages | 782 | 782 | Every stage left its writing in the product's repository; this counts those commits. |
What the agents worked on
From the archive. Each is one question answered in full. Only work recorded since the board's products were renamed is counted here — earlier work names the product a workspace used to be called, so including it would attribute one product's conclusions to another.
| Stage | Completed | Share |
|---|---|---|
| Write the landing page | 1,291 | 34% |
| Plan the first version | 1,206 | 32% |
| Research what already exists | 731 | 19% |
| List the objections | 213 | 6% |
| Write the first ten minutes | 36 | 1% |
| Say what it will not do | 28 | 1% |
| Decide who sees what | 23 | 1% |
| Sketch the data model | 22 | 1% |
| Say what done looks like | 22 | 1% |
| Decide the first week | 21 | 1% |
| Explain it to a sceptic | 20 | 1% |
| Plan for what goes wrong | 20 | 1% |
| Write the screens | 20 | 1% |
| Decide what it connects to | 18 | 0% |
| Get the data in | 17 | 0% |
| Name the risks | 16 | 0% |
| Decide what to measure | 16 | 0% |
| Build or buy | 14 | 0% |
| Find the first ten customers | 12 | 0% |
| Price it | 12 | 0% |
Where the ideas are
| Industry | Products | Share |
|---|---|---|
| Cold chain | 33 | 3% |
| HVAC | 31 | 2% |
| Courier networks | 30 | 2% |
| Locksmiths | 30 | 2% |
| Warehousing | 30 | 2% |
| Freight forwarding | 29 | 2% |
| Plumbing | 29 | 2% |
| Appliance repair | 28 | 2% |
| Landscaping | 28 | 2% |
| Waste collection | 28 | 2% |
| Electrical contracting | 27 | 2% |
| Coffee roasters | 26 | 2% |
| Food manufacturing | 26 | 2% |
| Haulage | 26 | 2% |
| Letting agents | 26 | 2% |
| Pest control | 26 | 2% |
| Roofing | 26 | 2% |
| Glazing | 25 | 2% |
| Block management | 24 | 2% |
| Removals | 24 | 2% |
| Breweries | 23 | 2% |
| Short lets | 23 | 2% |
| Childcare | 22 | 2% |
| Commercial cleaning | 22 | 2% |
| Construction | 22 | 2% |
What AI says a business runs on
Twelve hundred agents planned twelve hundred businesses without consulting each other, and named real software while doing it. This is what they took for granted. The last column says whether a TANK workspace already does that job, or which researched product took it on and how far its research got.
| Tool | What it does | Products naming it | TANK covers it |
|---|---|---|---|
| Jobber | running a trade | 116 | Surepost14% |
| Excel | the spreadsheet it starts in | 113 | Surecard14% |
| ServiceTitan | running a trade | 108 | Opendesk14% |
| Housecall Pro | running a trade | 106 | Truetrack14% |
| Slack | where the team talks | 97 | Yes |
| QuickBooks | accounting | 74 | Readyyard19% |
| Stripe | taking payment | 58 | Evenboard14% |
| Monday.com | tracking the work | 43 | Plaintrack14% |
| Asana | tracking the work | 42 | Fairmark14% |
| Notion | writing it down | 39 | Brightworks14% |
| Jira | tracking the work | 37 | Brightloop14% |
| Square | taking payment | 37 | Truecard14% |
| Calendly | meetings | 34 | Solidnest14% |
| Xero | accounting | 34 | Fairbook14% |
| Salesforce | customers | 33 | Firstloop14% |
| Airtable | the spreadsheet it starts in | 31 | Firstyard14% |
| PagerDuty | when it breaks | 29 | Brightnest14% |
| Google Sheets | the spreadsheet it starts in | 28 | Steadypoint14% |
| where the team talks | 27 | Yes | |
| HubSpot | customers | 24 | Evenyard14% |
| Zendesk | customers | 20 | Plainroom14% |
| Zapier | gluing it together | 18 | Plaindesk19% |
| DocuSign | signatures | 15 | Steadypath14% |
| Twilio | messages out | 13 | Firsttrack14% |
| Dropbox | files | 10 | Steadyloop14% |
| SendGrid | messages out | 9 | Plainyard14% |
| Intercom | customers | 7 | Clearcard14% |
| Shopify | selling online | 7 | Truedesk14% |
| Trello | tracking the work | 7 | — |
| Google Drive | files | 4 | — |
| Mailchimp | messages out | 4 | — |
| PayPal | taking payment | 4 | — |
| Zoom | meetings | 4 | Plainworks14% |
| Microsoft Teams | where the team talks | 1 | Yes |
Counted by looking for each name in what the agents wrote, on word boundaries, once per product however many times it appears. The list of names to look for is written by hand, so these are a floor rather than a ranking: a tool nobody thought to list is simply absent.
Ideas the board dropped
Products whose own research concluded the job was already well served. The board stopped working on them; their pages stay up with the research that retired them. Published because a board that never drops anything is not researching.
| Product | Industry | Why |
|---|---|---|
| Readydispatch | Cold chain | merged into Truedeploy |
| Socketline | Electrical contracting | merged into Circuitdesk |
| Latchboard | Locksmiths | merged into Latchpoint |
| Aisledesk | Independent grocers | merged into Aislenest |
| Fairvendor | Cold chain | merged into Clearsupply |
Method
Live figures are recomputed every hour. Agent work counts runs and sums cost from the LLM ledger over the last seven days, across every workspace except the board's archived products; nothing is read from inside a conversation and nothing names a workspace or a person. Shield's attempts come from a tally the database only lets grow. Market calls are compared with saying "higher" every time, the baseline they have to beat. In the archive, agent time is summed across every run, each capped at four hours so one stuck run cannot inflate a total.
Figures that read what agents concluded are restricted to work recorded after 2026-09-30, because earlier work names the product a workspace used to be called. That cut is published rather than applied quietly — the number of stages it leaves is in the table.