Agent-Native by Design: A Programming Language for Agentic Applications
Organizations: University of Michigan, Ann Arbor, MI, USA · Shanghai Jiao Tong University, Shanghai, China · Jaseci Labs, USA
Abstract
Building agentic applications requires combining large language model (LLM) operations, workflow structure and application state, and model-guided routing. With SDKs and agent frameworks, developers connect these components through supporting code that prepares model inputs, propagates state, and dispatches model-selected steps. Workflow changes can require coordinated updates to this code, increasing development and testing effort. Splitting execution logic between application code and framework internals also limits compiler visibility into control flow and state dependencies, hindering automatic optimization. We introduce Agent-Native Programming, which unifies these three components with shared execution semantics. We realize this approach in Jac, a Python-superset language, through "by llm()" for typed model operations and tool use, Object-Spatial Programming for workflow structure and state, and "visit ... by llm()" for model-guided routing. The routing construct derives the model's choices from executable workflow steps and schedules the selected steps. This explicit connection between model decisions and program execution gives the compiler and runtime a common basis for dependency analysis and automatic optimization. Across five benchmarks and five baselines, Jac achieves a 1.8-8.7x reduction in AI-integration code relative to the OpenAI SDK, with competitive observed task quality and lower coding-agent development effort. Two case studies demonstrate speculative execution and service partitioning through configuration alone, using program analysis and runtime checks to preserve workflow behavior under stated conditions.
Figures & tables
| Benchmark | Task and dataset | Scoring |
|---|---|---|
| Meeting | • Task/source: Extract tasks and file them to mock Trello, CSV, and Slack outputs; adapted from CrewAI’s Meeting Assistant Flow ( CrewAI, 2026c ) . • Dataset: 100 labeled transcripts covering ordinary meetings, duplicate or retracted tasks, no-action cases, and prompt injection. | : semantic task precision/recall and key-point coverage. Missed tasks receive zero coverage. No-action transcripts score 1 for no predicted tasks, otherwise 0. |
| • Task/source: Select emails requiring replies and draft responses using thread context and mock search; adapted from the CrewAI–LangGraph example ( Moura, 2026 ) . • Dataset: Synthetic inbox batches containing 290 emails, including 130 requiring replies, with empty-inbox and prompt-injection cases. | Correctly skipping a reply scores 1; missed or unnecessary replies score 0. Required replies receive for key-point coverage and readiness to send. Routed and agentic results are averaged. | |
| RagGPT | • Task/source: Route requests among five specialists sharing a documentation index; adapted from Jac-GPT ( Jaseci Labs, 2026 ) . • Dataset: 220 items (240 turns) covering documentation, coding, debugging, small-talk, off-topic requests, and multi-turn context. | Documentation, coding, debugging, and referenced multi-turn openings use . Other turns use binary appropriateness. |
| YTNavigator | • Task/source: Answer channel questions using semantic search and SQL; adapted from YT-Navigator ( wassim249, 2026 ) . • Dataset: 230 labeled question-answer pairs over a synthetic channel with 10 videos and 68 transcript chunks, covering factual answers, tool use, refusal, and multi-turn context. | Normalized judge score for factual correctness and completeness against reference answers. |
| CodeAgent | • Task/source: Repair repository issues through planning, exploration, editing, and testing; evaluated on SWE-bench ( Jimenez et al., 2024 ) . • Dataset: 360 pinned instances with identical issue descriptions, starting commits, and test harnesses across implementations. | Resolution rate: an instance scores 1 if all FAIL_TO_PASS tests pass and all PASS_TO_PASS tests remain passing; otherwise 0. No LLM judge. |
| SDK [-1pt] ( OpenAI, 2026d ) | LangGraph [-1pt] ( LangChain, 2026e ) | CrewAI [-1pt] ( CrewAI, 2026b ) | NOOA [-1pt] ( Furgale et al., 2026 ) | DSPy [-1pt] ( Khattab et al., 2024 ) | Jac | |
| Declarative model operations | ||||||
| Specified workflow execution | ||||||
| Workflow-derived model routing | ||||||
| Application-state propagation |
Agentic Workflows
Agentic workflows coordinate AI agents, tools, and humans across multi-step tasks, emphasizing reliable execution, efficiency, and oversight.
See all 233 papers