# Agent economics: evidence audit for the Knauf opening Research and retrieval date: 22 October 2026. Public primary sources only. This is a scoped claim audit, applying the local research-audit skill. It is not a reproduction of the experiments. Exact source locations, measurement boundaries and approved language are in [the claim ledger](../../../../data/evidence-upgrade/2026-10-22-evidence-upgrade/agents/2026-10-22-agent-economics-claims.json). ## Editorial judgement The strongest opening argument is that **AI economics are determined by the workflow, and additional spending must earn its place through accepted outcomes**. There is evidence of both useful extra effort and unnecessary effort. A universal forecast that AI unit costs will rise would overstate the evidence. Use at most two evidence moments from this audit. If another research stream supplies a credible adoption or expenditure trend, use only one of these in the opening and keep the other beside the optimisation section. 1. **Architecture changes resource demand.** Anthropic reports approximately 4× tokens for agents and 15× for multi-agent systems relative to chats in its own data. Both compare with chats. Label the metric “tokens”, the source and the limited scope on the slide. These are not bill multipliers. [Anthropic, June 2025](https://www.anthropic.com/research/multi-agent-research-system) 2. **More effort does not guarantee a better result.** HAL reports equal or lower accuracy at higher reasoning effort in 21 of 36 matched comparisons. Use “comparisons”, not individual tasks. Do not shorten this to “58% of AI spending is wasted”. [HAL, Figure 3 and §4.1](https://arxiv.org/html/2510.11977v1#S4) The first can be a sparse three-bar chart, indexed to chat = 1, with 1 / ~4 / ~15. The second can be 36 equal marks with 21 highlighted and a concise explanatory line. Neither needs a provider catalogue or a wall of statistics. Avoid presenting the two charts as one causal experiment: their methods and populations differ. ## Speaker bridge Suggested Geoffrey wording, approximately 30 seconds: > A token is getting cheaper in many settings. But the thing we deploy is increasingly a workflow: it searches, calls tools, checks and sometimes retries. That changes how much work a request sets in motion. The leadership question becomes: what does an accepted business result cost, and where is extra effort worth paying for? This is our inference from the evidence, not a quotation or a forecast. It leads directly into the existing accepted-result denominator and the three linked records: bill, workflow run and outcome. ## Counterevidence belongs in the spoken explanation Anthropic also reports a substantial gain on its internal research evaluation from multi-agent work. Its comparison is not a quality-controlled cost benchmark. The lesson is to use additional agents where their marginal value justifies them, not to ban them. HAL likewise finds that more tokens and accuracy are positively associated on several benchmark families. Its strongest lesson is workload-specific evaluation. The precise scopes and caveats are recorded in the ledger. ## Useful backup evidence The ledger includes three measured examples: context editing, variability across repeated coding-agent runs and cost-aware prompt optimisation. Each measures a particular system and workload. They support the practical options already in the deck; they do not establish expected Knauf savings. The later Anthropic architecture guidance is included as a fourth, alternative backup with a different comparator, not as confirmation of the original chart. For source quality, independent benchmark papers give more methodological detail than vendor engineering posts. The vendor posts are useful observed examples, with undisclosed sample sizes. The coding-cost paper is a recent working paper with a transparent benchmark setup; its extreme cross-workload headline should not be used. ## Claims deliberately rejected or held back - **“Agents cost 15× more than single agents.”** Incorrect comparator for the June 2025 source. - **“AI-agent bills will rise 15×.”** No aggregate-spend or price forecast follows from token observations. - **“High reasoning makes most tasks worse.”** HAL's denominator is 36 model–scaffold–benchmark comparisons; some ties are included. - **“Agents need 1,000× the tokens for the same task.”** The recent coding study compares agentic repository issues with separate reasoning/chat datasets. Workload differences invalidate this interpretation. - **“A coding-agent run usually varies by 30×.”** The coding paper's headline is an extreme; its repeated-run summary is approximately 2×. It retains complete conversation history, so other context policies can differ. - **“Cheaper list prices lose in 32% of enterprise selections.”** A May 2026 price-reversal preprint reports 106/336 model-pair/dataset comparisons, using input-plus-output list price as the ranking proxy and explicitly not controlling for quality. It also contains an inconsistent “one in five” prose summary. Useful corroboration of mechanism, insufficient basis for this general enterprise claim. [Version 2, §§2–3](https://arxiv.org/html/2603.23971v2) - **Old reliability results as current capability estimates.** The original τ-bench results remain methodological context, not a current 2026 capability claim. Repeated-run reliability is also different from finding one success after retries. [Original τ-bench](https://arxiv.org/abs/2406.12045) ## Practical measurement implication Record token categories and their prices separately, then reconcile with invoices. Include every call in a run, failed attempts, paid tools and review time. Summarise accepted-outcome cost and the distribution of run costs, rather than only the mean token count. Set actual runtime limits in the orchestration layer; a monthly budget notification and a model's self-estimated budget are not enforcement. These are design recommendations, not measured effect sizes from this audit. No verbatim source excerpts are stored. No source webpage, paper image or chart was copied into presentation assets. The numerical charts recommended above should be redrawn from the small data arrays in the ledger and visibly attributed.