RFC #3: Improve API Realism for Tau3

[victorb-sierra]

Problem
Current APIs in Tau2 tasks may not reflect real-world usage or complexity, limiting the benchmark’s usefulness for evaluating agents in realistic scenarios.

Proposal

• P0: Update APIs to be more realistic and reflective of tools used in production.
• P1: Introduce failure modes or constraints that agents must handle.

Discussion Points

• Which APIs should be prioritized for realism improvements?
• How should failure modes be designed to balance realism with testability?
• Any compatibility concerns with existing tasks or evaluations?
• Is it easy to agree on what "realistic" APIs are?

Next Steps (after consensus)

• Break proposed changes into actionable Feature/Task issues.
• Assign issues to the v3.0 milestone and v3 branch for PRs.