Answers from the documentation; checking which model is on duty
Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
Stanford updates AI-native engineering course with focus on coding age - OpenSmartRoute
Introduction - what the course teaches and why it was created
Stanford's CS146S course now focuses on teaching students how to build and use coding agents. These agents are AI systems that can write, review, and explain code. The course was created to prepare students for a rapidly changing software development landscape. It reflects recent advances in autonomous coding, where AI tools help programmers work faster and more effectively.
The course aims to teach students principles of AI-driven coding practices. It covers architecture, open-source collaboration, and how to create reliable AI agents. The goal is to help future engineers understand how to integrate AI into their workflows. This course was designed because coding agents are becoming a core part of software engineering. It replaces traditional methods by emphasizing automation and AI collaboration.
History of CS146S - how the course evolved over time
CS146S started with nine assignments and a final project. Its first run was in fall 2025. The course was built around exploring how AI agents can assist in coding tasks. Initially, students used AI to review code, evaluate feedback, and perform other programming activities. The focus was on exposing students to the broad potential of AI in software development.
Over time, the course evolved as AI coding tools improved. The first version aimed to give students a survey of different AI applications. It introduced them to various coding agents for tasks like review, specification, and site reliability engineering. As AI tools advanced, the course shifted toward teaching core principles and best practices. The second iteration, in fall 2026, emphasizes architecture, context engineering, and open-source contributions. This reflects the growing maturity of AI coding systems.
Recent changes in AI coding practices - new principles and tools
Recent developments have led to new principles for building AI coding systems. One key idea is careful context engineering. This involves designing how AI agents understand and handle complex codebases. It helps ensure AI systems produce accurate and safe results. Another principle is focusing on architecture. Developers now think more about how AI agents fit into larger workflows and systems.
November 6, 2026 — ARC Prize tests Inkling 2 Small, and Thinking Machines explains how it caught early checkpoints bypassing guardrails.
Tools have also advanced. Students now learn to build custom workspaces where AI agents can run and test code. These workspaces allow AI systems to reproduce failures and validate patches. They are tailored to each project’s needs, making AI integration more reliable. The emphasis is on creating systems that are both autonomous and controllable, reducing risks and improving outcomes.
Open-source contributions - how students engage with real projects
Students are now required to contribute to open-source projects as part of the course. This gives them real-world experience working with actual codebases. They learn to ask maintainers for feedback and review their patches. This process teaches them how to collaborate with experienced developers and understand project standards.
Contributing to open-source also helps students see how AI agents can assist in practical tasks. For example, agents can implement features or review pull requests. Students learn to ask agents for explanations and to generate visual walkthroughs of code changes. This hands-on approach makes the learning process more relevant and prepares students for modern software workflows.
Monaco’s engineering practices - building custom workspaces for AI code
At Monaco, the engineering team has developed custom workspaces for AI coding agents. These workspaces are environments where AI systems can run, reproduce failures, and test patches. They are tailored to Monaco’s cloud setup, ensuring smooth integration and testing. The team believes that building their own workspaces helps them control the AI development process.
Having dedicated workspaces allows Monaco to evaluate AI tools more effectively. Engineers can ask agents to investigate issues or explain code changes. They can also test new models and see how they perform in real scenarios. This approach helps Monaco improve its AI systems and integrate them into production workflows more safely.
Use cases of coding agents - from code review to explanation
Coding agents are now used for many tasks beyond simple code generation. They assist in code review by analyzing pull requests and providing feedback. Agents can also generate explanations for code changes, making it easier for teams to understand why certain decisions were made. This helps improve transparency and collaboration.
Another common use case is debugging. Agents can investigate failures, reproduce bugs, and suggest fixes. They are also used to implement new features by translating specifications into code. These agents can handle complex tasks, such as managing dependencies or optimizing performance. Their versatility makes them valuable tools in modern software development.
Principles of architecture and context engineering - handling complex codebases
Designing AI systems for large codebases requires careful architecture. Context engineering involves structuring how AI agents access and interpret code. Developers create authority maps that guide agents to relevant parts of the code. This helps ensure that AI systems understand the current behavior, past decisions, and project goals.
Handling complex codebases also means designing workflows where AI agents can learn from outcomes. If an agent’s suggestion fails, it adjusts its approach over time. This feedback loop improves accuracy and safety. Proper architecture and context engineering are essential for making AI systems reliable in production environments.
Impacts on software development - faster, more autonomous workflows
The integration of AI coding agents is changing how software is developed. Teams can now automate many routine tasks, such as code review and bug investigation. This speeds up workflows and reduces manual effort. AI agents also enable more autonomous development, where systems can generate and test code with minimal human input.
These changes lead to faster iteration cycles and quicker deployment. Teams can handle more complex projects because AI helps manage dependencies and optimize performance. Overall, AI-driven workflows make software development more efficient and scalable. They also allow engineers to focus on higher-level design and innovation.
How OpenSmartRoute helps - routing requests for AI development and testing
OpenSmartRoute is an open-source router designed for AI requests. It directs each request to the best-fit model, agent, or tool from a predefined catalog. The router scores candidates based on quality, cost, speed, and safety. Teams can set weights for each factor depending on their needs.
Using OpenSmartRoute helps teams optimize AI workflows. It ensures requests are handled by the most suitable models, reducing costs and improving results. The router also learns from outcomes, giving more traffic to successful models. This makes AI development and testing more efficient and reliable. Teams can track costs and performance through built-in metrics and safeguards.
How it compares - what existed before, what this changes and what stays the same
Before the use of coding agents, software development relied heavily on manual coding, reviews, and debugging by humans. Engineers would write code, submit pull requests, and wait for reviewers to examine changes. This process could be slow and prone to human error, especially in large projects.
Traditional code review involved human reviewers examining code changes, providing feedback, and approving or rejecting pull requests. This process was often time-consuming and depended on the availability and expertise of reviewers. It also lacked automation and real-time feedback.
The new approach introduces AI agents that can handle many tasks automatically. These agents can review code, suggest patches, investigate failures, and even contribute to open-source projects. They can perform tasks in a fraction of the time required by humans, enabling faster iteration cycles.
What stays the same is the core role of human engineers in decision-making. AI agents assist but do not replace the need for human oversight. Engineers still decide which changes to accept, review the AI’s suggestions, and maintain the project’s overall direction.
This shift changes the workflow by making it more continuous and automated. Instead of waiting days for reviews, teams can get instant feedback from agents. It also expands what engineers can handle, allowing them to focus on higher-level design rather than routine tasks.
Questions this leaves open - what the source does not say and how a reader can check it
The source does not specify the exact technical details of how the AI agents are built or trained. It does not mention the underlying models, algorithms, or datasets used. Readers cannot determine the specific architecture or training process from the provided information.
It also does not clarify the limits of the current AI agents. For example, how well do they handle complex or novel problems? Are there cases where they fail or produce unsafe suggestions? Without evaluation results or benchmarks, it is hard to assess their reliability.
The source does not explain how teams measure the success of these AI tools. Do they track improvements in review speed, code quality, or bug rates? How do they compare to traditional workflows in quantitative terms? Readers may need to look for case studies or performance reports to answer these questions.
Another open question concerns the ethical and safety considerations. How do teams ensure that AI agents do not introduce security vulnerabilities or biased code? The source mentions permissions and dependencies but does not detail safety protocols or oversight mechanisms.
To check these aspects, a reader can explore additional documentation or research papers from the teams involved. They can look for published benchmarks, user feedback, or evaluations of the AI agents in real-world settings. Engaging with the open-source community or attending conferences like AIE Code can also provide deeper insights.
Finally, the source does not specify how widely adopted these practices are outside of Stanford and Monaco. Are other companies or open-source projects using similar AI-driven workflows? What are the barriers to adoption? Investigating industry reports or talking to practitioners can help answer these questions.
