{
  "slug": "za-co-placi-sie-w-projekcie-agentowym",
  "ranking_id": "Z11",
  "edition": "1.0",
  "date": "2026-09-26",
  "series": {
    "pl": "Lider i jego warsztat",
    "en": "The leader and their workspace"
  },
  "number": 3,
  "title": {
    "pl": "Za co płaci się w projekcie budowanym z agentami",
    "en": "What you pay for in an agent-built project"
  },
  "intro": {
    "pl": "Ludzie, narzędzia, odbiór i utrzymanie. Neutralny rachunek pokazuje, dlaczego abonament modelu nie jest kosztem całego rezultatu.",
    "en": "People, tools, acceptance and maintenance. A neutral calculation shows why a model subscription is not the cost of the complete result."
  },
  "source_dates": [
    "2026-09-16",
    "2026-05-11",
    "2026-02-24"
  ],
  "checked_at": "2026-09-26",
  "sources": [
    {
      "title": "Koszt wykonania w podejściu agentowym",
      "access": "internal",
      "date": "2026-09-16",
      "scope": "Full method adapted; private estimates and vendor pricing excluded"
    },
    {
      "title": "METR: Self-Reported Impact of Early-2026 AI",
      "url": "https://metr.org/blog/2026-05-11-ai-usage-survey/",
      "accessed_at": "2026-09-26"
    },
    {
      "title": "METR: Changing Developer Productivity Experiment Design",
      "url": "https://metr.org/blog/2026-02-24-uplift-update/",
      "accessed_at": "2026-09-26"
    }
  ],
  "evidence_scope": {
    "pl": "Fikcyjne założenia, sprawdzona arytmetyka. Bez oferty, pomiaru oszczędności, wyceny projektu ani spółki.",
    "en": "Fictional assumptions, checked arithmetic. No quotation, savings measurement, project estimate or company valuation."
  },
  "update_trigger": {
    "pl": "Błąd rachunku, zmiana założeń lub istotna korekta źródła.",
    "en": "An arithmetic error, changed assumption or material source correction."
  },
  "ai": {
    "authoring": "Codex",
    "source_selection": "Claude Code + Codex",
    "human_review": "post-publication withdrawal by Lech R. Rustecki",
    "independent_model_review": false
  },
  "interactive": "budget-diagram",
  "schema_version": 1,
  "language": "en",
  "url": "https://www.l00p.ai/en/resources/series/za-co-placi-sie-w-projekcie-agentowym/",
  "markdown_sha256": "6b60dccbe36a726c394880caaa9aafedd8154be69986227e66300a3838c7d626",
  "sections": [
    {
      "id": "rozdzial-01",
      "title": "01 / The model bill is part of the bill",
      "url": "https://www.l00p.ai/en/resources/series/za-co-placi-sie-w-projekcie-agentowym/#rozdzial-01",
      "markdown": "An agent produced code in an hour. Did the project cost one hour of work and a few model calls? Someone still had to describe the need, connect data, check the result, fix errors and decide what happens after a failure. Those activities do not disappear from a budget just because they are not code writing.\n\nThe starting point is an internal delivery-cost analysis dated 16 September 2026. We retain its method of separating effort, but do not publish estimates for specific projects. The example below is entirely fictional. It is an arithmetic exercise, not a quotation, a market price list or a savings measurement. Estimating software delivery effort is also not valuing a company."
    },
    {
      "id": "rozdzial-02",
      "title": "02 / First define what must pass acceptance",
      "url": "https://www.l00p.ai/en/resources/series/za-co-placi-sie-w-projekcie-agentowym/#rozdzial-02",
      "markdown": "Instead of “build an AI system”, name one process: receive material, prepare a result, make it reviewable and save it in an agreed place. Define inputs, outputs, failure cases and the reviewer. Establish which integrations already work and which still need investigation.\n\nSeparate reproducing a known function from discovering a product. A working reference, field definitions and available data change the scope. File or code-line counts alone do not tell you the cost of bringing a process to acceptance. [Conducting agent work](/en/resources/series/lider-jako-dyrygent-pracy-agentowej/) helps connect an assignment, its result and completion evidence.\n\nName exclusions in the estimate: archive migration, hardware, licences, long-term maintenance or additional support coverage. An out-of-scope item does not become free. It needs a separate estimate or a deliberate decision to defer it."
    },
    {
      "id": "rozdzial-03",
      "title": "03 / A neutral example: a small delivery stage",
      "url": "https://www.l00p.ai/en/resources/series/za-co-placi-sie-w-projekcie-agentowym/#rozdzial-03",
      "markdown": "Assume 40 hours of active human work and an illustrative rate of PLN 150 per hour. Using one rate for every role simplifies the exercise; it is not a pricing recommendation. The hours cover:\n\n- 8 h defining the process and acceptance criteria;\n- 12 h implementation and directing agents;\n- 8 h integration with data and tools;\n- 8 h tests, acceptance and corrections;\n- 4 h deployment and handover.\n\nThe total is 40 h, so human work amounts to PLN 6,000. Add an assumed PLN 800 budget for tools and trial service use. The planning base is PLN 6,800. A contingency of 20% of that base is PLN 1,360, bringing the plan including contingency to PLN 8,160. Every figure is an assumption or arithmetic based on those assumptions.\n\n![Human work PLN 6000, tools 800, contingency 1360: plan 8160. Excluding unused contingency: PLN 6800 divided by 10 or 8 accepted results gives PLN 680 or 850.](https://www.l00p.ai/wydawnictwo/za-co-placi-sie-w-projekcie-agentowym/budzet-en-v1.svg)\n\nOriginal numerical diagram, Codex / L00P.AI. Segment widths represent amounts. All data are fictional; contingency is not expenditure incurred.\n\n[Download assumptions and calculations as CSV](/wydawnictwo/za-co-placi-sie-w-projekcie-agentowym/przyklad.csv). The file contains the same figures with Polish/English descriptions. It contains no project data or current service prices. Changes to the rate, scope or risk require recalculation."
    },
    {
      "id": "rozdzial-04",
      "title": "04 / Contingency, expenditure and price differ",
      "url": "https://www.l00p.ai/en/resources/series/za-co-placi-sie-w-projekcie-agentowym/#rozdzial-04",
      "markdown": "Contingency makes room for uncertainty. It is not evidence of money spent. In this example, we do not add it to unit cost as though it has already been used. For time-based billing, track budget consumption and the remaining buffer. For a fixed outcome price, treatment of risk depends on the agreed terms of working together.\n\nDo not charge the same subscription twice: once to the project as a whole and again to every task. Define how shared tools are allocated. If the owner does the work, their time may not be a cash payment to a contractor, but it remains an input. Show cash expenditure separately from the assigned value of work instead of blending them into a seemingly comparable price.\n\nAn hour of several agents running in parallel is not automatically several hours of human labour. Record active human time and service consumption separately. Waiting for a result may extend elapsed delivery time even when it is not all active work."
    },
    {
      "id": "rozdzial-05",
      "title": "05 / Divide by accepted results",
      "url": "https://www.l00p.ai/en/resources/series/za-co-placi-sie-w-projekcie-agentowym/#rozdzial-05",
      "markdown": "Further assume that the fictional stage cost excluding unused contingency was PLN 6,800. With ten comparable, accepted results, unit cost is PLN 680. With eight, it is PLN 850. Total cost has not changed; the denominator has. These are two scenarios, not observed outcomes from our publishing operation.\n\nDo not count a rejected answer as a finished result. Include correction work in the numerator and acceptance in the denominator. If no result has been accepted, cost per accepted result is undefined, not zero. Compare similar tasks with similar quality requirements; one difficult process is not interchangeable with ten easy drafts."
    },
    {
      "id": "rozdzial-06",
      "title": "06 / Speed is not a discount",
      "url": "https://www.l00p.ai/en/resources/series/za-co-placi-sie-w-projekcie-agentowym/#rozdzial-06",
      "markdown": "[METR's May 2026 survey](https://metr.org/blog/2026-05-11-ai-usage-survey/) distinguishes reported speed gains from the value of work produced. The authors describe self-reports and sampling limitations. These findings do not automatically determine the price of a particular implementation. The [February study-design update](https://metr.org/blog/2026-02-24-uplift-update/) also describes difficulties in measuring changing AI work practices.\n\nOur practical proposal is to choose a representative, bounded stage and measure active human time, tool expenditure, corrections and accepted results. Record tool versions, task scope and date. Then revise the estimate's assumptions. Do not promise a percentage saving based only on code generation speed."
    },
    {
      "id": "rozdzial-07",
      "title": "07 / The budget continues after deployment",
      "url": "https://www.l00p.ai/en/resources/series/za-co-placi-sie-w-projekcie-agentowym/#rozdzial-07",
      "markdown": "After acceptance come updates, monitoring, backups, incident handling and service use by the product. In a separate fictional scenario, maintenance costs PLN 600 per month. Twelve months total PLN 7,200. Added to the delivery plan including contingency, this gives PLN 15,360. That is a sum of planned items, not historical expenditure or a complete ownership cost for every situation. Taxes, hardware and unlisted licences are outside this exercise.\n\nFor a conversation about your own project, one process, an acceptance criterion and a list of existing integrations are enough. We can help distinguish effort, cash, risk and maintenance. Private costs and implementation details can be discussed without publishing them in an example."
    },
    {
      "id": "rozdzial-08",
      "title": "08 / Sources, arithmetic and corrections",
      "url": "https://www.l00p.ai/en/resources/series/za-co-placi-sie-w-projekcie-agentowym/#rozdzial-08",
      "markdown": "The full internal analysis dated 16 September was read. METR sources were checked on 26 September 2026. Private ranges, supplier quotations and historical model prices were not carried over. The example's arithmetic was checked locally; no pilot measuring team speed or estimate of an actual project was performed.\n\nCodex prepared the Polish/English text, CSV and deterministic diagram. Review was by the author, without an independent second-model assessment. No third-party illustrations were used. A person may withdraw the material after publication. An arithmetic error, changed assumption or material source correction triggers another review; updating the date requires an actual check."
    }
  ],
  "media": [
    {
      "url": "https://www.l00p.ai/wydawnictwo/za-co-placi-sie-w-projekcie-agentowym/budzet-en-v1.svg",
      "kind": "diagram_in_code",
      "depicts_real_measurement": false
    }
  ]
}