PROJECT INTEL
Zurvan
An open-source, goal-driven Python framework for building production AI agents.
- ACTIVE SINCE:
- 2025 — present
- STATUS:
- ACTIVE
- FIREPOWER
- 9/10
- ARMOR
- 9/10
- SPEED
- 7/10
- SPECIAL
- 8/10
Zurvan — named after the Zoroastrian deity of time — is an open-source framework for building AI agents that survive contact with production. Agents are modelled on a composable GAME loop (Goals, Actions, Memory, Environment), and behaviour is injected through Capability hooks rather than subclassing — so the agent core stays minimal and testable.
Tool inputs are Pydantic-validated, and when a model produces something invalid, the agent self-corrects from the validation errors instead of crashing. A LiteLLM abstraction keeps the whole framework provider-agnostic: swap Claude, OpenAI or Gemini without touching agent code. Published on PyPI.
BATTLE RECORD
- Composable GAME-loop architecture: Goals, Actions, Memory, Environment
- Capability hooks instead of subclassing — minimal, testable agent core
- Pydantic-validated tool inputs with LLM self-correction on validation failure
- Provider-agnostic via LiteLLM; published on PyPI
TECH
- Python
- GAME-loop architecture
- Pydantic
- LiteLLM
- Multi-agent
- PyPI