OpenAI described a research collaboration with Ironclad to test agents using software for contracting. Tasks include configuring forms, agreements and approval rules, where completing clicks is not enough: the final workflow must respect business rules.
The team defined 11 tasks with multiple success criteria. OpenAI reports a higher score for GPT-6 Astra than GPT-5.6 Sol in this research evaluation, but the result is specific to these tasks and does not remove the need for human review.
If you build an agent, define testable success criteria, exercise exceptions and check the final outcome rather than only the sequence of actions.
Want to apply what you read?
Projects ↗