AgentBarrier 1.0.0: keep a human in the loop when AI agents want to take action
AgentBarrier now lets you approve or block what your AI agents do, with dashboards, Slack, and secure record-keeping built in for production use.
AI agents that can call tools or take actions on their own are powerful, but that power needs a check. AgentBarrier 1.0.0 introduces a core enforcement system that decides, based on rules you set, whether an agent's action should proceed, and it keeps a record of that decision. This works whether your agents are built with OpenAI Agents, LangGraph, PydanticAI, or Google's Agent Development Kit, since each of these frameworks now has its own connector to AgentBarrier's approval system. There's also a gateway for agents that use the MCP protocol, so those tool calls are checked and validated too.
To act on these approvals, you get a secure web dashboard where pending requests show up in real time, complete with an audit trail of what happened and when. If your team prefers to work from Slack, you can review and approve or deny requests right from there instead of switching tools. Approval events can also trigger signed webhooks, so other systems can be notified reliably and the notifications can be verified as authentic.
For teams running this in production, AgentBarrier now supports PostgreSQL as a storage option alongside SQLite, so audit logs and approval records can scale with your setup. Access to the dashboard and approval workflows is controlled by role-based permissions, so different team members can have different levels of authority. Operators also get controls to pause, resume, or limit how much an agent runtime can do, giving a way to step in if something needs to stop.
This release also fixes a privilege configuration issue in the PostgreSQL storage backend and strengthens how the system recovers from errors and validates policy rules, so the enforcement behavior stays consistent and predictable.
With this release, approving or blocking an AI agent's actions is no longer something you have to build yourself. You can see what agents are asking to do, decide from a dashboard or Slack, and trust that the record of those decisions is secure and auditable, whether you're running one agent framework or several.
How to use it
Connect the adapter for your agent framework (OpenAI Agents, LangGraph, PydanticAI, or Google ADK) so its tool calls route through AgentBarrier. From there, review and act on pending approvals using the web dashboard or directly in Slack, and set up webhooks if you want other systems notified when approval decisions are made.