Self-adjudication trap asks how mechanical gates can protect agent code from inference waste

Generating code by agents has become low-cost, but announcing task completion has turned into the most costly dilemma in software engineering. In automated development environments, practical experience shows that the agent that writes the code remains a weak judge, unqualified to assess the maturity of its work, leading systems into a trap of false confidence or recurring code loops.

Faced with this obstacle, a research team led by Si Hwa Liu, Tony Nuttaski and Song Han proposed the “Humanize” framework, a multi-agent coordination schema based on what they described as adjudication engineering. The idea places decisive, strict decisions at the boundaries separating planning, execution, review and learning phases, rather than leaving the decision to a single language model’s judgment.

The workflow starts with human approval of the contract, then a building-execution agent handles the task over defined rounds, while the final decision and termination are delegated to a review agent belonging to a completely different model provider.Language models do not steer tasks between these roles; instead, mandatory code hooks enforce 72 strict mechanical gates.When the process is viewed as a Markov chain across software-repository states, the joint pull from two different models ensures that no software bug slips through unless both models simultaneously overlook it at the same moment.

The system was evaluated across 68 releases over 108 days, garnering 1,468 stars on GitHub, with its results documented in 118 publicly available anatomical examinations of real workflow loops. Practical applications included migrating the gem5 platform build system, which comprises 567 files, and deploying core-design agents that secured the top three positions in all full-agent tracks of the Flash Infer competition at MLSys 2026. The framework also achieved perfect scores in the International Olympiads for Informatics, Mathematics, Physics and Biology, won a gold medal in Chemistry with a score of 418.5 out of 437, topped the LinEval math leaderboard with 251 out of 303, outperforming professional mathematicians, and recorded a perfect 672 on the Batnam Bench test.

However, the anatomical examinations revealed a core flaw no less serious than coding errors:Identifying the stopping point remains the major weakness in agent operation.While independent review succeeded in spotting unsupported achievement claims, reports showed that two-thirds of software-work rounds were consumed after execution had already been accepted, field observational data that highlights operational waste.

This shift carries a decisive implication for technology managers and engineering teams in Gulf banks and software firms in Egypt and the Levant. The rush toward programming agents without mechanical verification gates means draining computing budgets on self-referential inference loops with no real return. The required skill for local developers is no longer crafting textual prompts, but constructing strict hand-off gates and separating the review entity from the building entity to enforce quality and prevent waste.