Developers working with large language models (LLMs) have increasingly adopted the DSML tool calls parser convention in recent system updates, according to technical documentation released this year. This structured wrapper, which maps executable calls to XML-based parsing semantics, was implemented to standardize agent and tool-calling pipelines across specific model formats, sources confirmed.
This convention maps executable calls to XML-based parsing semantics, specifically employing tags such as `<|DSML|invoke name="...">` and `<|DSML|parameter name="...">`, according to documentation cited by vLLM and other technical sources released in 2024. The parser treats the DSML wrapper as a shell alias, while the internal parsing logic remains XML-driven, officials said.
The DSML tool calls parser convention, identified by the markup `<|DSML|tool_calls>`, is primarily a structured wrapper used in large language model (LLM) systems to standardize the execution of external tool or API calls, sources confirmed.
The preferred DSML markup form begins with `<|DSML|tool_calls>` and is followed by nested `<|DSML|invoke>` and `<|DSML|parameter>` tags, according to a parser source referenced in recent technical literature. A legacy XML form using ``, ``, and `` tags remains compatible with current implementations, ensuring some backward compatibility. However, older tags such as ``, ``, ``, and `` are explicitly treated as plain text rather than executable instructions, the source added. Malformed wrappers released by the parser are also output as plain text without execution, maintaining strict input validation.
Other tool-call syntaxes, including ``, `tool_use`, `antml` variants, and pure JSON fragments labeled `tool_calls`, are not recognized as executable calls in the DSML parser implementation, sources confirmed. This strict separation reduces ambiguity for runtime parsers tasked with identifying the start and end of tool invocation messages. The parser’s behavior is implementation-specific but consistently enforces recognition only of the DSML shell and its XML-equivalent forms as valid executable calls.
The tool-calling workflow enabled by DSML conventions allows LLM agents to output structured data that invokes external APIs or functions. Typically, a tool definition includes a name, description, and parameters, which the model generates in a tool-call message. This message is then executed by the server, with the results fed back into the model for further synthesis, according to research and talk materials from 2023 and 2024. This process is fundamental for agents requiring external actions such as retrieving weather data or querying databases.
Recent research, including a 2026 paper titled *W&D: Scaling Parallel Tool Calling for Efficient Deep Research Agents*, highlights advances in parallel and delegated tool use. These approaches enable models to generate code that chains multiple function calls and leverage sandboxed code execution tools that act as callers for other tools, reducing the need for the LLM itself to orchestrate every step. A YouTube talk from late 2023 further detailed these developments, describing multi-step orchestration and delegated execution as emerging standards.
Additionally, tool retrieval and context-window optimization have become key focuses in tool-calling systems. Anthropic’s introduction of a tool search mechanism, as described in the same 2023 talk, reduces the need to load extensive tool schemas into the model’s context window. This approach reportedly uses only about 500 tokens in some scenarios, saving up to 80% of context window usage. Deferred loading, where tools are dynamically retrieved rather than statically injected, is part of this trend toward more efficient and scalable tool-calling pipelines.
It is important to distinguish that the acronym DSML traditionally stands for domain-specific modeling language in academic literature. Formal modeling papers from 2022 to 2024 define DSML as a four-tuple comprising a domain and its interpretations, often used for software agents based on belief–desire–intention frameworks. The parser-related usage of DSML as a tool-call shell alias represents a different operational meaning specific to LLM runtime environments, according to sources analyzing both academic and technical documents.
These developments reflect ongoing efforts within the AI research community to standardize and optimize how language models interact with external tools. The DSML tool calls parser convention, while not a universal protocol, serves as a practical standard within certain model ecosystems to ensure consistent and unambiguous tool invocation. Further updates and refinements to these conventions are expected as LLM architectures and agent workflows evolve.
