Agent Kernel

Version: 1.0.4-stable

AGenNext Kernel is the execution engine for AgentNext. It receives pre-validated requests from AGenNext Runner, executes the requested work, persists execution state, and returns results and traces.

Kernel does not own user-facing configuration or boundary policy enforcement. Those responsibilities are split across the platform:

Platform → defines policies, guardrails, tools, memory config, and framework options
Runner   → validates requests, loads config, and enforces policies at the runtime boundary
Kernel   → executes pre-validated work, persists state, and returns results/traces

What This Solves

AI agents need a reliable execution substrate after policy and guardrail checks have already passed.

AGenNext Kernel provides that substrate by handling:

  • Controlled execution of pre-validated tasks
  • Tenant-scoped execution context
  • Persistent memory and execution state
  • Traceable results for audit and observability
  • Adapter-backed calls to external systems such as GitHub, Slack, and Jira

Kernel Responsibilities

Kernel is responsible for execution-time behavior:

  • Run validated agent tasks from Runner
  • Execute tool calls and framework operations
  • Apply tenant context to execution, memory, and trace operations
  • Persist execution results, decisions, memory updates, and traces
  • Return structured results to Runner

Kernel assumes Runner has already performed input validation, policy checks, guardrail enforcement, and configuration loading.


Non-Responsibilities

Kernel should not duplicate responsibilities owned by Platform or Runner.

Kernel does not:

  • Define OPA/Rego policies
  • Own user-facing guardrail configuration
  • Decide whether an unvalidated request is allowed
  • Act as the primary policy-enforcement boundary
  • Replace Runner framework, tool, memory, guardrail, or trace adapters

For policy authoring and runtime enforcement, use Platform and Runner respectively.


Core Capabilities

  • Execution API for pre-validated requests
  • Streaming execution support
  • Tenant-scoped memory operations
  • Trace and audit retrieval
  • External-system execution adapters
  • Structured result handoff back to Runner

API Surface

The Kernel API is expected to expose execution-focused endpoints:

  • POST /execute — run a pre-validated task with platform-provided config
  • POST /stream — stream execution events and partial results
  • POST /memory — perform tenant-scoped memory operations
  • GET /trace/{id} — retrieve execution traces

Requests should include tenant context and any pre-validated execution metadata supplied by Runner.


Quickstart

docker compose up --build -d

Related Repositories

  • AGenNext Platform — defines policies, guardrails, tools, framework options, and user-facing configuration
  • AGenNext Runner — enforces policies, validates requests, loads Platform config, and invokes Kernel
  • AGenNext Kernel — executes pre-validated work and returns results/traces

Docs

See the /docs folder for API and architecture details.


© 2026 Agent Kernel Team


Discover more from AGenNext

Subscribe to get the latest posts sent to your email.

Comments

Leave a Reply