Agent-net Open Sources Webagent For Guarded AI Business Agents

Agent-net, a team focused on building an agent-to-agent marketplace for AI interactions, has launched Webagent, an open-source harness designed to enable public-facing business agents. This new tool allows businesses to transform their websites into AI agents that can interact with other agents. Instead of requiring extensive orchestration code, businesses can configure an agent by providing a declarative JSON specification and selecting providers for nine pluggable slots, then running the `webagent serve` command. The Webagent harness is distributed under the Apache 2.0 license, builds successfully, and currently supports live Slack, WhatsApp, and HTTP agents powered by MCP tools. However, it is still in its initial v0 release, with features such as a browser action provider, OAuth-gated MCP, OTel export, and the AgentNet identity and billing layer still under development.

The Webagent architecture is built using the Go programming language. Each agent comprises a "Brain," which consists of a Large Language Model (LLM) combined with specific instructions, and a set of "slots." These slots are defined in the `core/` directory and represent Service Provider Interfaces (SPIs). The `spi/` directory contains a registry of providers for each slot, with a designated default provider. The built-in slots and their default providers include Retrieval (live, keyword, hybrid - live), Memory (session), Guardrail (basic, off - basic), Channel (a2a, web, slack, whatsapp, with a telegram stub - a2a), Secrets (env, file, static - env), Presenter (text, terminal, web - text), Model (echo, openrouter, gateway - echo), Action (none, demo, mcp - none), and Observability (none, log, memory - none).

This design caters to three distinct user groups under a single contract: businesses that configure agents by selecting from predefined options, businesses that extend functionality by registering custom providers, and partner companies that develop and ship adapter modules. Crucially, the core of Webagent is not intended for forking. To ensure quality and compatibility, every provider must successfully pass its respective slot's conformance suite to achieve certification.

A fundamental architectural decision within Webagent is the implementation of `action.Guard`. This mechanism ensures that every tool available to the agent, whether provided by the action provider or injected by the host system, is wrapped by a selected guardrail. This guardrail is executed before the action itself, preventing the LLM from bypassing safety protocols. The project's design philosophy, as outlined in its `DESIGN.md` document, centers on this robust safety enforcement, prioritizing code-based security over prompt-based limitations.