Recoleta Item Note
Agentis – An AI-native programming language where the LLM is the stdlib
Agentis proposes a programming language for AI agents that treats the LLM directly as the “standard library” and represents code as a version-controlled binary hashed DAG rather than text files. Its goal is to make…
ai-native-languageagent-programmingllm-as-stdlibcontent-addressed-codeevolutionary-branching
Summary
Agentis proposes a programming language for AI agents that treats the LLM directly as the “standard library” and represents code as a version-controlled binary hashed DAG rather than text files. Its goal is to make programs naturally centered around prompts, validation, branch exploration, and controlled execution.
Problem
- Traditional programming languages treat LLMs as external APIs rather than language-level primitives, which makes building agent programs centered on reasoning, classification, and extraction relatively cumbersome.
- The text-file + conventional VCS model is not necessarily well suited to AI-generated/evolutionary code, making merge conflicts more likely and lacking native branching and tracing mechanisms centered on agent execution.
- Autonomous agents can easily produce runaway calls, unbounded resource consumption, or unsafe I/O, so budget constraints, sandboxing, and a verifiable execution model are needed.
Approach
- Design
promptas a language primitive: many operations traditionally handled by a stdlib are instead turned into direct requests to the LLM, with support for typed outputs such asprompt(...) -> list<string>or structured types. - Use a
validatemechanism to enforce constraints on LLM outputs, such as confidence thresholds, thereby incorporating generated results into program logic and failure-handling flows. - Introduce a Cognitive Budget (CB) / fuel mechanism to limit operation costs, preventing agents from running indefinitely and encouraging more efficient prompt design.
- Use
explorefor evolutionary branching: execution can fork, successful results form new branches, and failed branches are discarded, making it suitable for search-style problem solving. - Store code as a SHA-256 content-addressed AST/DAG and fuse it with a built-in VCS; also provide sandboxed file I/O and domain-whitelist network access to improve security.
Results
- The text does not provide formal experiments, benchmarks, or peer-reviewed quantitative results, so its performance gains relative to existing languages/frameworks cannot be verified.
- It provides one operability claim: the first run of
agentis go examples/fast-demo.agproduces output in about 3–8 seconds, but does not specify hardware, model, task, or comparison baseline. - It presents several concrete feature claims: support for typed prompt outputs, validation rules, evolutionary branching, content-addressed code storage, a merge-conflict-free design, sandboxed I/O, and network whitelisting.
- At the toolchain level, it lists commands and workflows such as
init,go,commit,run,branch,switch, andlog, indicating that this is not just a language concept but also includes execution and version-management implementation. - The implementation-complexity claim is strong: it asserts “Zero bloat,” is based on Rust, and mentions dependencies only on sha2 and ureq, but this is an engineering description rather than an outcome metric.
Link
Built with Recoleta
Run your own research radar
Turn arXiv, Hacker News, OpenReview, Hugging Face Daily Papers, and RSS into local Markdown, Obsidian notes, Telegram digests, and a public site.