Agentic Kernel

A microkernel agent runtime: a minimal, governed "plan → authorize → execute → observe" loop with everything pluggable behind injection seams. Available as two feature-paired SDKs:

• TypeScript — @agentic-kernel/* monorepo (core + adapters)
• Python — agentic-kernel (single distribution; agentic_kernel.core + adapters)

Both SDKs are at v0.6.0 and kept at feature parity.

Why a kernel?

Most agent frameworks bake the harness — model calls, memory, policy, retries, logging — directly into the loop. Agentic Kernel keeps the loop minimal and governed, and composes the harness at well-defined seams instead. That buys:

• Reproducibility — every run is an immutable, replayable execution log.
• Governance — policy, approvals, budgets, and capability/credential checks are first-class seams, not scattered conditionals.
• Bounded autonomy — a hard iteration ceiling and stall→reflect recovery are built into the kernel; no run is unbounded.
• Portability — swap the model, store, memory, or observability backend by swapping an adapter; the kernel never changes.
• Testability — contract test suites (conformance) prove any adapter is a safe substitute.

Start here

• Architecture — the run loop, data model, and the injection seams.
• Getting started — first agent in either SDK.
• Consuming packages — install from the registry.
• Guides: Memory · On-device models · Multi-agent orchestration · Token usage
• Manuals: Developer guide · Comprehensive manual
• Development records — real-LLM evaluation & stress campaigns.

Status

Public and Apache-2.0: TypeScript on npmjs (@agentic-kernel/*), Python on PyPI (agentic-kernel-runtime; import package agentic_kernel). See Installation.