Show HN: Tadpole – an LLM agent built from scratch then post-trained on 4B model (github.com)

Tadpole, a newly released LLM agent, is notable for being developed entirely from scratch in about 3,500 lines of plain Python code, without relying on an existing agent framework. This tool-using agent can interact with files, execute shell commands, and perform web searches through local or OpenAI-compatible APIs. Tadpole's construction entails a meticulous, stage-by-stage development process that captures incremental improvements and validations, significantly enhancing its capabilities. Notably, post-training elevated the Qwen3-4B model's performance from a pass rate of 0.419 to 0.774 on tasks it had not previously encountered. This initiative is significant for the AI/ML community as it emphasizes transparency in the development of LLM agents, offering comprehensive tutorials and test measurements along each phase. Key technical elements include the agent's ability to maintain memory across sessions, utilize context budgeting, and execute background tasks efficiently. Furthermore, the project serves as a platform for self-improvement by allowing real-time feedback to refine prompts and tools. With the freedom to run in environments such as Telegram or locally, Tadpole presents powerful implications for both developers and users interested in flexible, capable AI agent systems.