Most languages treat AI agents as an external add-on: you call an API, parse the response, and glue everything together yourself. Jac, a Python-like language from Prof. Jason Mars of the University of Michigan and Jaseci Labs, takes a different position — AI agents are first-class citizens of the language, built into how programs are written rather than bolted on afterwards. It was demonstrated at All Things Open in Raleigh, North Carolina, in October 2026.
Jac is a Python superset, so Python developers start from familiar ground, and it runs on the Jaseci runtime. It integrates the Python and TypeScript runtimes, compiles to Python bytecode, JavaScript and native machine code, uses LLVM for Rust-style borrow checking, and can import C and C++ libraries. Developers get access to both the PyPI and npm ecosystems. Its distinctive programming model is graph-based Object-Spatial Programming: data lives in nodes and edges, and "walkers" move through the graph carrying computation with them. Mars describes the language in his book "A Synechic and Topokinetic Programming Language" (Edition 1.1.0, July 2026, published on Zenodo). The language is open source at github.com/jaseci-labs/jac.
The headline research claim comes from a November 2026 paper (arXiv:2611.21298), which reports a 1.8–8.7x reduction in AI-integration code versus the OpenAI SDK across five benchmarks. That claim needs its caveat, stated plainly. Harness Hub's adversarial review (19 November 2026) flags that the comparison rests on author-written baselines and author-supplied Jac skills; under fixed designs, Claude Code shows no token advantage over the plain SDK; and the speculative-execution latency gains appear on only 3 of the 5 benchmarks. The paper's numbers are a starting point for discussion, not a settled verdict.
One more practical note: Java is not supported. Jac's world is Python, TypeScript and C/C++ — Java developers should not expect interop there.
Key facts
| Item | Details |
|---|---|
| Language | Jac — Python-like / Python-superset language |
| From | Prof. Jason Mars (University of Michigan, Jaseci Labs), with the Jaseci runtime |
| Core idea | AI agents as first-class citizens |
| Demonstrated at | All Things Open, Raleigh, NC, October 2026 |
| Runtimes | Integrates Python + TypeScript runtimes |
| Compiles to | Python bytecode, JavaScript, native machine code |
| Safety | LLVM for Rust-style borrow checking |
| Interop | Imports C/C++ libraries; PyPI + npm ecosystems |
| Programming model | Graph-based Object-Spatial Programming — nodes, edges, "walkers" |
| Book | "A Synechic and Topokinetic Programming Language" by Jason Mars (Edition 1.1.0, July 2026, Zenodo) |
| Repository | github.com/jaseci-labs/jac (open source) |
| Paper claim | arXiv:2611.21298 (Nov 2026): 1.8–8.7x reduction in AI-integration code vs OpenAI SDK across 5 benchmarks |
| Caveat | Claim rests on author-written baselines and author-supplied Jac skills (Harness Hub adversarial review, 19 Nov 2026); Claude Code shows no token advantage over plain SDK under fixed designs; speculative-execution latency gains on only 3 of 5 benchmarks |
| Not supported | Java |
What this means
Jac is best understood as a bet on how software will be written next. If AI agents become a normal part of programs rather than an external service, then a language that treats agents as first-class citizens — with graph-shaped data and walkers that carry work through it — could make agent-heavy applications much simpler to express. The graph model is genuinely different from the object-and-function style most developers know, and it is worth studying even if you never adopt Jac itself.
But the honest reading of the evidence matters here. A 1.8–8.7x reduction in integration code would be a big deal — if it held up under independent testing. The Harness Hub review shows it does not yet: author-written baselines, author-supplied skills, no token advantage for Claude Code under fixed designs, and latency gains on only three of five benchmarks. That does not make Jac a bad language; it makes the headline number a claim to verify, not a fact to repeat. Students should learn from the pattern: whenever a paper's authors write both the tool and the test, read the comparison with extra care.
For Python developers, Jac is approachable — a superset means your existing knowledge transfers — and the open-source repository is the right place to start experimenting.
A note on verification
The language description, author, runtime, demonstration, compilation targets, borrow checking, library interop, ecosystems, programming model, book details and repository are as verified in the re-verify returns (21 November 2026). The 1.8–8.7x figure is the November 2026 paper's own reported result (arXiv:2611.21298); the caveat comes from Harness Hub's adversarial review of 19 November 2026. Java is not supported. This post will be updated if independent evaluations of Jac appear.
Source: Jaseci Labs / Prof. Jason Mars; All Things Open (Raleigh, NC, Oct 2026); arXiv:2611.21298; Harness Hub adversarial review, 19 Nov 2026. Last checked: 21 November 2026.
FAQs
What is Jac?
Jac is a Python-like, Python-superset programming language from Prof. Jason Mars (University of Michigan, Jaseci Labs) with the Jaseci runtime. AI agents are first-class citizens of the language, and it uses graph-based Object-Spatial Programming with nodes, edges and "walkers".
What can Jac compile to, and what does it work with?
Jac compiles to Python bytecode, JavaScript and native machine code. It integrates the Python and TypeScript runtimes, uses LLVM for Rust-style borrow checking, imports C/C++ libraries, and draws on the PyPI and npm ecosystems.
Is the 1.8–8.7x code-reduction claim proven?
No. The November 2026 paper (arXiv:2611.21298) reports 1.8–8.7x less AI-integration code than the OpenAI SDK across five benchmarks, but Harness Hub's adversarial review (19 Nov 2026) flags that the claim rests on author-written baselines and author-supplied Jac skills. Claude Code shows no token advantage over the plain SDK under fixed designs, and speculative-execution latency gains appear on only 3 of 5 benchmarks.
Does Jac support Java?
No. Jac works with Python, TypeScript and C/C++ — Java is not supported.
Comments (0)
Share your question or experience about this syllabus. Your comment will appear after review.
No comments yet — be the first to share your thoughts!