Programming agents finish their tasks before completing them: how DeReact architecture untangles the trust dilemma

The most dangerous thing a software engineer faces when building autonomous systems is not the smart agent stopping work, but its confident claim that it has completed the task successfully while leaving the code broken in the runtime environment. This disconnect between claimed completion and actual reality represents today the major knot that stalls the transition of agents from experiments to sensitive production architectures. Within the widely adopted React architecture, a single language-model policy handles proposing the next step, interacting with the code environment, and deciding when the task ends. This merging of execution and oversight lets errors slip in without deterrent, and allows the agent to finish its run without possessing evidence that it has fulfilled its assignment.

Faced with this structural dilemma, researchers Ajay Vohra, Tao Chen, Neti Narayan, and Karen Chang presented in a research paper deposited in early October 2026 on the arXiv platform an alternative architecture they called “DeReact” that dismantles this closed loop.The new architecture removes the central model’s monopoly on judgment by delegating oversight to two independent external gateways: a software reviewer that checks the safety of proposed actions before execution, and a context manager that reconstructs the system state from the actual runtime environment and formally certifies task completion.This shift turns the agent from an entity that validates its own assumptions into a system governed by external verification mechanisms that accept a claim of completion only after its evidence is examined.

Benchmark results the team ran on the GAI​An and SWE Bench Verified platforms revealed a highly significant technical pattern. The biggest gains did not go to the leading, high-cost models, but were concentrated in the less capable models; the “Qwen-3-Coder-480B” model achieved a jump of 6.5 to 7.0 points on the accuracy metric from the first attempt, and the “Claude Sonnet 4.5” model recorded improvements ranging from 4.2 to 5.2 points. In contrast, those gains diminished with the “Claude Opus 4.5” model, whose success rates remained close to the traditional architecture, but with a decisive difference manifested in producing fully documented, constraint-compliant work traces, meaning the system preferred deliberation and scrutiny over hasty termination.

This shift in design philosophy reshuffles spending priorities for technical engineering teams in the Gulf and Egypt, especially in banks and government agencies that race against time to automate their software processes and manage their data.Betting on acquiring the heaviest, most expensive language models is not the only path to reliability; data confirm that adjusting the architecture and distributing authority gives open-source, medium-sized models a competitive efficiency that matches large models at a far lower inference cost.Moreover, imposing external inspection gateways aligns with compliance and governance requirements emphasized by regional regulatory frameworks, where every step and every termination decision becomes subject to programmatic documentation that can be audited and held accountable instead of relying on a black box.

The most salient lesson we observed in this experiment is that external oversight gateways succeed when model failures are common and tangible, and when the gateway policy itself has sufficient auditing competence.If you lead a team building smart agents today, your engineering step begins by stripping the model of authority to endorse its work;Thus, task completion is not a claim generated by the model, but a certification endorsed by the runtime environment.