Praxa

Praxa is an agent execution framework (harness) proposed in an arXiv paper, aimed at making the chain from agent authorization to effect transparent and testable; its core contribution is an evidence-bound architecture that makes the state transition from authorization to effect explicit and testable.

The Praxa paper distinguishes five types of claims in agent systems: proposal, authority, dispatch, verified external effect, and serving promotion.

Praxa makes each authorization-to-effect transition explicit and testable through five stages: deterministic admission, brokered execution, external read-back, reconciliation, and post-audit promotion.

Deterministic admission refers to the system determining, according to explicit and predictable rules, whether an action is permitted to enter the execution queue before any action is executed, rather than relying on the model's own judgment.

Brokered execution means that actions are not triggered directly by the agent but are dispatched through an intermediate broker layer responsible for isolation, logging, and rate limiting.

External read-back refers to actively querying the external system for state after execution to confirm whether an action truly took effect, rather than inferring the result solely from the return value of the execution call.

On a fixed revision of the repository, Praxa passed 1,027/1,027 unit tests and 89/89 Workerd tests, instrumented all 363 expected source files, and met four coverage lower bounds, though raw per-test records and independent reproduction are not yet available.

In the Terminal-Bench Core 0.1.1 pilot across 12 selected tasks, both the baseline group and the reliability-layer group each passed 17/36 under strict trials, with the reliability layer consuming 37.49% more input tokens and 50.73% more output tokens.

The paper explicitly acknowledges that the Terminal-Bench pilot does not support a superiority conclusion, as the reliability layer delivered no better success rate at higher cost.

In a two-phase coordination-proxy development comparison, both the baseline and the source-author-written candidate completed 180/180 trials; the candidate used 37.11% fewer tokens, reduced estimated endpoint cost by 33.84%, and decreased steps by 11.63%, but improvement in quality, latency, or production-environment behavior cannot be demonstrated.

The Praxa paper explicitly states that current evidence cannot demonstrate adversarial security, production-environment safety, general expert-level superiority, autonomous recursive optimization capability, or user benefit.

The evidence-binding architecture requires every state transition in the system to be accompanied by verifiable evidence rather than relying solely on logical inference, which is in line with the concepts of idempotency verification and transaction logging in traditional software engineering.

Adding reliability layers incurs significant token and cost overhead without necessarily yielding immediate improvements in success rates, and the benefits of auditability and reliability often come at the expense of efficiency.