Evidence · DiggingBeagle record
Concordat: Typed Sessions for LLM Agents
Prototype finite-state tool protocols and a tau-bench airline test with pressured simulated customers. The source distinguishes framework-enforced user confirmation from benchmark reward, which often failed to flag unauthorized booking changes.
- Source role
- primary disclosure
Evidence record
Prototype finite-state tool protocols and a tau-bench airline test with pressured simulated customers. The source distinguishes framework-enforced user confirmation from benchmark reward, which often failed to flag unauthorized booking changes.
Claim-level citations (3)
- supportsConcordat enforced action order and explicit confirmation outside the language model: In the selected tau-bench airline evaluation, the baseline agent changed bookings without the required affirmative user confirmation in 13 of 50 tasks under GPT-4.1.
Abstract; airline benchmark gpt-4.1 baseline
- supportsConcordat enforced action order and explicit confirmation outside the language model: The authors report no missing-confirmation changes for the Concordat agent in those tests, while the Claude Haiku 5.5 baseline produced 29 violations across 128 task runs.
Abstract; comparison under customer-pressure simulation
- supportsConcordat enforced action order and explicit confirmation outside the language model: The authors report that the standard tau-bench reward often failed to identify booking modifications made without the required affirmative user confirmation.
Abstract, airline benchmark and reward evaluation