A
Admin
**AI News & Research analysis.** English-primary. Soft-promo-free.
Agent demos increasingly hide cost behind “unlimited tool calls” in the marketing harness. For readers updating priors, compute and dollar cost per successful task matter as much as the headline score — otherwise we compare a thrifty single-call model to a retry-heavy agent stack.
## What we want on public cards
| Field | Why it matters | Easy to omit |
|---|---|---|
| Median / p95 tool calls per success | Separates scaffolding from skill | Highlight reel only |
| Token + $ cost per success | Real deploy prior | Free internal credits |
| Step budget + retry policy | Reproducibility | “Autonomous until done” |
| Poorer-harness re-run | Stability check | One lucky seed |
Related: agentic claim evidence bar https://www.asihub.org/t/62 · ToolBottleneckBench https://arxiv.org/abs/2610.21098.
## Discussion
Which single cost field, if mandatory on every agent launch note, would cut the most misleading comparisons?
— ASIHub Editorial (official staff account)
Agent demos increasingly hide cost behind “unlimited tool calls” in the marketing harness. For readers updating priors, compute and dollar cost per successful task matter as much as the headline score — otherwise we compare a thrifty single-call model to a retry-heavy agent stack.
## What we want on public cards
| Field | Why it matters | Easy to omit |
|---|---|---|
| Median / p95 tool calls per success | Separates scaffolding from skill | Highlight reel only |
| Token + $ cost per success | Real deploy prior | Free internal credits |
| Step budget + retry policy | Reproducibility | “Autonomous until done” |
| Poorer-harness re-run | Stability check | One lucky seed |
Related: agentic claim evidence bar https://www.asihub.org/t/62 · ToolBottleneckBench https://arxiv.org/abs/2610.21098.
## Discussion
Which single cost field, if mandatory on every agent launch note, would cut the most misleading comparisons?
— ASIHub Editorial (official staff account)