Agentend

Agentend is a Python framework for building agent as a backend systems, where the agent handles the application’s backend role rather than relying on traditional routes, controllers, and related infrastructure. It is aimed at agent-powered products and supports code-defined capabilities instead of endpoint-based design.

The framework’s documented flow starts with a user intent, described as natural language rather than a structured API call. A kernel classifies and routes the intent, with the page describing this step as using a small model with under 10 ms latency. It then hydrates memory through a context bus with five tiers of progressive loading, and workers execute tasks while events stream in real time through the AG-UI protocol. The same page also highlights typed worker slots, DAG workflows with parallel steps and human-in-the-loop interrupts, a dual-layer semantic cache based on hash and vector similarity, PALADIN injection defense, per-tenant token budget enforcement, eight hook points with YAML manifests, and pip-installable extensions.

Agentend includes a chat-to-build interface for creating capabilities without code, and the example on the site shows a capability defined in Python with a decorator, worker names, domain context, and a tool function. The architecture section names AG-UI, MCP, and A2A as part of a “protocol triangle,” and the observability and persistence notes mention OpenTelemetry traces and metrics, plus SQLAlchemy 2.0 async with multi-tenant row-level security. Installation commands shown on the page use pip and a local serve workflow, and the code sample imports Agentend, Capability, and tool from the package.

The site describes the project as production-ready and says it is licensed under AGPL-3.0-or-later. It also names Cloud, Documentation, Quickstart, Fleet, Memory, Events, and Community in its navigation, and the page title identifies it as a Python framework.