Skip to content
Agentic Patterns

Agentic Patterns

Composable primitives for building LLM agents. Frozen, immutable atoms compose upward into roles and agents, executed by a runtime with typed events, gates, and exporters.
Get startedGitHub

The shape of the framework

Agent = Role × Background × Awareness × Mission
Role  = Persona + Judgments + Capabilities + Responsibilities
Capability = Toolbox + Manual + Playbook

Everything composes upward through layers — atoms → protocols → molecules → rendering → organisms in @agentic-patterns/core, then events → gates → runner → transport → runtime → workflows → conversation → exporters → presets in @agentic-patterns/runtime. Core never imports runtime.

Install

Terminal window
bun add @agentic-patterns/core @agentic-patterns/runtime

The server (@agentic-patterns/server) and CLI (@agentic-patterns/cli) version in lockstep with the runtime; core floats independently.

Where to start

  • Getting started — zero to a running agent, with or without an API key.
  • Guides — authoring toolboxes and plays; the runner & provider strategy.
  • Memory — cross-session memory: the store, the recall surface, and the evolution cookbook.
  • Reference — the SSE event catalog (generated from the runtime manifest) and playground event persistence.
  • Architecture Decisions — the ADR trail, from the constellation dashboard to compositional memory.
  • Design notes — dated design and planning documents, kept honest and clearly separated from shipped API.

Live API surfaces ship with the server itself: Scalar API docs at /docs and an llms.txt at /llms.txt on any running instance.