Saved in:
| Main Author: | |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2026
|
| Online Access: | https://doi.org/10.5281/zenodo.19647938 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- <div> <div>Autonomous AI agent platforms --- systems that decompose natural-language instructions into executable plans, invoke tools, and verify results without human intervention --- have emerged as a critical frontier in software engineering automation. Existing commercial platforms are cloud-dependent, vendor-locked, and lack extensibility, while open-source alternatives (OpenHands, Suna) offer narrow tool sets without multi-provider model routing or domain-specific skill injection. We present Kabab, an open-source hybrid architecture that combines Rust for performance-critical subsystems (Docker sandbox management, parallel skill indexing, lock-free LLM API proxy, browser WASM compute at 173KB) with TypeScript for a agent pipeline orchestrating 25 models across 11 providers. Kabab introduces two novel contributions: (1) a Universal Skill Protocol that auto-loads domain-specific expertise from SKILL.md files based on task classification across 10 task types, and (2) Connector-Augmented Execution that dynamically registers 185 SaaS API actions across 21 services as LLM-callable tools at runtime, enabling an agent to autonomously clone a repository, fix a bug, push code, and create a pull request from a single instruction. We evaluate Kabab on real software engineering tasks using both local (Qwen 3.5 35B via Ollama) and cloud (Groq Llama 3.3 70B, Claude Opus) models, demonstrating that hybrid local-cloud agent execution is feasible, cost-effective, and produces functional multi-file applications with database schemas, API endpoints, and styled frontends in a single autonomous session. Kabab is open-source under Apache 2.0.</div> <div> </div> </div>