How Microsoft Foundry's Code Interpreter Actually Runs Python Inside a Sandbox

·8 views

Microsoft Foundry's Agent Service includes a Code Interpreter tool that spins up an isolated Python container to perform deterministic computations, rather than relying on a language model to reason through calculations. The system addresses a core weakness of large language models, which are prone to errors in arithmetic, string manipulation, and large dataset operations. When an agent invokes the tool, Foundry lazily provisions a sandboxed container, stages files through Azure Storage, executes the code, and returns verified output with no network access to external infrastructure. Three key components govern the system: the tool definition, the container environment, and the session, which is a billable, time-limited runtime unit. Understanding the distinction between auto-managed and explicitly managed containers, as well as the difference between prompt agents and hosted agents, is critical for developers deploying the feature at production scale.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

AI Writes Frontend Code, But Developers Still Need Core Skills to Thrive

AI tools can now generate functional frontend components in seconds, prompting debate about what skills developers still need to master. Experts argue that foundational knowledge of HTML semantics, CSS layout, JavaScript behavior, and browser internals remains critical for evaluating and correcting AI-generated code. Real-world projects demand far more than writing code — they require understanding UX, architecture, accessibility, performance, and debugging. AI may suggest multiple fixes for a problem, but a developer who understands concepts like CORS or render pipelines can diagnose issues far more efficiently. Rather than replacing core knowledge, AI tools make it more important for developers to reason critically about the code they accept and ship.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why AI Chatbots Confidently Choose the Wrong Tool When Options Are Ambiguous

When AI systems are connected to multiple external tools, a critical failure point emerges in how the model decides which tool to use for a given request. Unlike obvious requests, real-world user queries are often ambiguous enough to appear valid for more than one tool, causing inconsistent routing decisions. The consequences are frequently subtle — the wrong tool may still return a response, just a less accurate or less useful one — making such errors difficult to detect through standard monitoring. The stakes rise sharply when routing confusion involves tools with side effects, such as payment processing or booking changes, where an incorrect selection can trigger irreversible actions. Developers are urged to treat tool-selection logic as a critical system component rather than a trivial default, particularly when read-only and write-capable tools coexist.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

LLMs Belong Inside Recommendation Stacks, Not on Top of Them

Recent 2025-2026 research papers, including RecoChain and RRCM, argue that large language models should function within specific layers of a recommendation system rather than replacing the entire pipeline. A standard production recommender operates across four layers: data and features, candidate retrieval, ranking, and feedback, with LLMs contributing only to retrieval and semantic feature extraction. Strict latency requirements, typically around 100 milliseconds end-to-end, make it technically infeasible to place a generative model call in the critical request path for large catalogs. Engineers are advised to precompute or cache LLM outputs, such as item embeddings and query understanding, and always maintain a deterministic fallback when model calls fail. A commonly overlooked data practice is logging impressions alongside clicks, which is essential for computing click-through rates, correcting position bias, and enabling offline ranker evaluation.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

IQueryable vs IEnumerable: How Your LINQ Choice Decides Where Filtering Happens

In C#, IEnumerable and IQueryable share identical LINQ syntax but execute filtering in fundamentally different places — memory versus database. IEnumerable processes data inside the application after fetching all rows, while IQueryable builds an expression tree that gets translated into SQL and executed at the database level. A common anti-pattern involves returning IEnumerable from a repository method, which silently forces the entire table into memory before any filtering occurs, causing severe performance issues at scale. Developers are advised to keep IQueryable as long as possible in the query chain and only switch to IEnumerable via AsEnumerable() when applying custom C# logic that cannot be translated to SQL. Calling ToList() or AsEnumerable() marks the boundary where database execution ends and in-memory processing begins.

0 comments Read more at DEV Community