Code Agent
Timeline (last 90 days)
The behavior of code agents is highly malleable, shaped not only by the task description itself but also by the physical structure of the repository and whether priorities are explicitly expressed.
Under otherwise equal conditions, three supportive local files yield a 96% adoption rate, whereas sibling-resource files achieve only a 4% adoption rate.
With the number of examples held fixed, merely changing where the files are placed can make the adoption rate jump from 0% to 100%.
When the task intent explicitly requires differing from existing code, in a joint task with two agents and two implementation pairings, code examples achieve only a 38/60 joint success rate.
Written priority rules achieve a 59/60 joint success rate in the joint task, outperforming the 38/60 of code examples.
The study covers four code agents and six implementation pairings, conducting a total of 5,625 trials and validating in four open-source repositories, involving three programming languages.
Different code agents exhibit highly consistent default implementation preferences when no precedent exists.
When a precedent conflicting with the default preference exists in the repository, agents follow the precedent almost without exception—in 360 relevant trials, 359 chose the precedent.
In real projects, coding agents still need version control, test coverage, and human code review as a safety net, because their self-reported "tests passed" does not equal quality assurance.
The speaker suggests that when teams introduce agent engineering, they first give the agent a small task with clear boundaries, provide architecture documents, repository structure, conventions, and constraints; require a plan before execution and review the plan; define acceptance criteria; enforce verification methods; and keep human checkpoints around security-sensitive or high-impact operations.
2 more timeline events
All Facts (13)
Top-tier code agents are already capable of autonomously resolving more than 50% of real GitHub issues.
70%Unverified代码智能体正在从交互式开发者工具演进为可编程的开发基础设施
70%UnverifiedThe behavior of code agents is highly malleable, shaped not only by the task description itself but also by the physical structure of the repository and whether priorities are explicitly expressed.
50%UnverifiedWith the number of examples held fixed, merely changing where the files are placed can make the adoption rate jump from 0% to 100%.
50%UnverifiedWhen the task intent explicitly requires differing from existing code, in a joint task with two agents and two implementation pairings, code examples achieve only a 38/60 joint success rate.
50%UnverifiedWritten priority rules achieve a 59/60 joint success rate in the joint task, outperforming the 38/60 of code examples.
50%UnverifiedThe study covers four code agents and six implementation pairings, conducting a total of 5,625 trials and validating in four open-source repositories, involving three programming languages.
50%UnverifiedDifferent code agents exhibit highly consistent default implementation preferences when no precedent exists.
50%UnverifiedWhen a precedent conflicting with the default preference exists in the repository, agents follow the precedent almost without exception—in 360 relevant trials, 359 chose the precedent.
50%UnverifiedUnder otherwise equal conditions, three supportive local files yield a 96% adoption rate, whereas sibling-resource files achieve only a 4% adoption rate.
50%UnverifiedIn real projects, coding agents still need version control, test coverage, and human code review as a safety net, because their self-reported "tests passed" does not equal quality assurance.
50%UnverifiedThe speaker suggests that when teams introduce agent engineering, they first give the agent a small task with clear boundaries, provide architecture documents, repository structure, conventions, and constraints; require a plan before execution and review the plan; define acceptance criteria; enforce verification methods; and keep human checkpoints around security-sensitive or high-impact operations.
50%UnverifiedThe working principle of a coding agent is essentially to manipulate the computer by writing code.
50%