Abstract
A serving system constructs and executes a directed acyclic graph (DAG) at runtime for each request based on an agent model’s tool-call decisions. The DAG is incrementally formed from base tools and intent-selected tools chosen using request intent, conversational context, and intermediate results. Runtime safety constraints are enforced, including a maximum path length, per-tool repetition bounds to avoid cycles, allowlist enforcement, and a token budget. A tool registry provides per-tool description, capability tags, and estimated token and latency metadata that may be used to filter or prioritize tool selection under remaining budgets. Execution may include multiple phases, with later phases conditionally executed based on predicates evaluated from earlier-phase results, such as running listwise optimization in one embodiment when a strong-candidate threshold is met.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Runtime-Constructed Execution Graphs with Agent-Directed Path Selection and Constraint Enforcement for Serving Systems", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10717