Estimating the maximum cost of an agent run
Model the spending limit of an agent workflow, including retries, with a worked example and an interactive calculator.
Method draft · Illustrative ceiling · Review pending
Research
Study plans, methods and findings from our work on AI systems.
Reviewing the version that will ship
Trace how review evidence and approval attach to the exact version being released.
Method draft · Public evidence extract pending
When the timestamps ran ahead
Five record pairs show what can be compared between an agent’s time estimates and repository timestamps.
Case draft · Event pairing review pending
Does the workflow work after the form saves?
A saved record failed to appear in the connected list in the Allegory benchmark. A second check exposed a sorting gap.
Engineering case summary · Synthetic workflow · Public evidence review pending
What makes an AI citation useful?
A 53-entry source register and 24 bilingual query concepts underpin separate checks for visibility and claim support.
Study plan summary · Query tests proposed · Public evidence review pending
When can an agent reuse what it remembers?
A proposed design for tracking a claim’s source, revision history and the permissions to act on it or share it.
Proposed method summary · Implementation and validation pending
How should a small team test a change?
A measurement review connects decisions to offline checks, matched comparisons and randomized trials.
Proposed method summary · Public evidence review pending
No research matches these filters.