Small models. Real workflows.
⁂asterism
A modular agent harness for SLMs.
Turn language models into agents that carry out workflows and automations. Orchestration, durable memory, tools and shared inference work together, with execution grounded in evidence.
↳ Four components, one harness
⁂ asngn
asterism engine
Coordinates workflows, tracks task state and applies execution policies. Checks outcomes against evidence to carry work through to completion.
View on GitHub ↗ 02 / MEMORY⁂ asper
asterism persistence
Preserves identity, context, project knowledge and execution evidence beyond a single context window. Combines exact events, semantic memory and checkpoints.
View on GitHub ↗ 03 / ACTION⁂ astools
asterism tools
Connects model intent to real operations through installable tool packages. Typed contracts, permissions and supervised execution make tools discoverable and controllable.
View on GitHub ↗ 04 / INFERENCE⁂ asmodel
asterism model runtime
Shares model resources across the harness. Manages memory budgets and reusable contexts, with a common API for generation, embeddings and explicit provider capabilities.
View on GitHub ↗↳ Working together
asngn coordinates the work, asper retains its context and evidence, astools carries out actions, and asmodel supplies shared inference.
Designed to be friendly to small language models and local inference, the harness also supports larger models and mixed local and remote configurations.