Agent‑net’s Webagent solves the biggest headache for businesses that want to put AI agents to work without writing custom orchestration code. Instead of wiring together LLMs, tools, memory, safety checks and channel adapters by hand, a team fills out a single JSON spec that picks one provider for each of nine plug‑in slots: retrieval, memory, guardrail, channel, secrets, presenter, model (brain), action and observability. Running webagent serve spins up a ready‑to‑talk agent that can speak over Slack, WhatsApp, HTTP or agent‑to‑agent channels immediately.
The core safety guarantee is built into the code: every tool call first passes through action.Guard, which runs the chosen guardrail before the action executes. The model cannot bypass this wrapper, so unsafe behaviour is prevented by architecture rather than by prompting. This gives teams confidence that their agents will not perform prohibited actions even if the underlying LLM tries to trick them.
What works today out of the box includes the echo, OpenRouter and gateway brains, the MCP action provider for connecting to existing MCP servers, live Slack and WhatsApp webhook handlers with signature verification and de‑duplication, HTTP agent‑to‑agent and web channels, and secret vaults that read from environment files or static values at build time. Observability emits per‑turn traces aligned with OpenTelemetry GenAI conventions, and an eval harness lets you run scenario tests.
What is still missing are browser‑based actions, OAuth‑guarded MCP servers, a full OpenTelemetry exporter, a Telegram adapter, partner‑supplied memory or guardrail plug‑ins, and the AgentNet identity and billing layer. Because the project is Apache 2.0 licensed and written in Go, teams can add their own providers by implementing the service‑provider interface and passing the conformance suite—no forking of the core is required.
For businesses that need a secure, observable, channel‑agnostic AI agent without the usual engineering overhead, Webagent offers a declarative, safe‑by‑design path from spec to running service in minutes.