Yoram Kornatzky

Nov 3

More operations. Tighter controls. Over the past few days, I expanded Agent Workspace, my Laravel platform for goal-defined agents, while tightening how agents use tools. 1. Expanded connectors. Added Buffer, Trello, and Cursor alongside the existing connectors. A broader set of operations gives m...Read more

More operations. Tighter controls. Over the past few days, I expanded Agent Workspace, my Laravel platform for goal-defined agents, while tightening how agents use tools. 1. Expanded connectors. Added Buffer, Trello, and Cursor alongside the existing connectors. A broader set of operations gives me more meaningful cases for studying multi-action prompts and workflows. 2. Preventing invented actions. Constrained agents to actions related to their prompt, rather than letting them invent additional tasks. Having access to a tool does not mean the agent should use it. 3. Approval for tools used by other tools. A tool calling another tool does not bypass approval. This includes subordinate reads needed to complete another action. 4. File operations. Added file operations, with human approval for write and destructive actions. Using the Laravel AI SDK's tool approval mechanism. 5. Web search. Added web search, with human approval, using the Laravel AI SDK. Screenshots show the approval request and the results after approval. 6. Schema validation. Added validation for tool inputs and outputs. Tool calls need to respect their contracts, not just produce plausible-looking data. The focus is not just adding capabilities. It is keeping control as the agent gains more ways to act. How do you handle approval when one tool needs to invoke another? #AgentWorkspace #Laravel #LaravelAI #AIAgents #HumanInTheLoop #ToolCalling #BuildInPublic

Oct 30

Connecting an agent to Gmail and GitHub is the easy part. The harder part is enforcing permissions and approvals across a workflow. In Agent Workspace, I have implemented separate approvals for write and destructive actions. Now I am investigating whether an agent performing several actions can by...Read more

Connecting an agent to Gmail and GitHub is the easy part. The harder part is enforcing permissions and approvals across a workflow. In Agent Workspace, I have implemented separate approvals for write and destructive actions. Now I am investigating whether an agent performing several actions can bypass that approval scheme, and how my implementation should use Laravel AI SDK tool approval. Example: matching Gmail messages with GitHub issues. Several tool calls, each with its own permission boundary. For those building tool-using agents: how do you enforce approvals across a chain of calls?

Oct 27

Agent Workspace is becoming my agent research lab. It is built with Laravel, the Laravel AI SDK, and Composio. The application now executes actions. Write or destructive operations are each approved separately. The last few days were research rather than new implementation. I separated the question...Read more

Agent Workspace is becoming my agent research lab. It is built with Laravel, the Laravel AI SDK, and Composio. The application now executes actions. Write or destructive operations are each approved separately. The last few days were research rather than new implementation. I separated the questions into three scopes: 1. An action. Permissions and approval before changing an external system. 2. A workflow. Several dependent steps, such as taking a Notion specification through to a GitHub PR.Multiple users. 3. Agents communicating on behalf of different people. The mechanism is still open. These are research directions, not three new milestones. The example of a workflow raised a concrete question. What if the process crashes after reading Notion but before creating a GitHub issue? Or GitHub times out after creating it? Approval does not answer whether the work can resume safely. A timeout does not establish whether the external action happened. I have not implemented that recovery yet. For now, the progress is identifying what to study in the working application, rather than turning every research topic into a feature. For those of you building agents that use tools, when did you move from a chat-driven agent loop to a durable workflow? And what forced the change?

Oct 22

Approving an agent’s goal should not approve every action it takes. Today, I added separate approvals for write and destructive actions in Agent Workspace. Built with Laravel, the Laravel AI SDK, and Composio. The existing approval lets the agent work toward a goal. Each write or destructive acti...Read more

Approving an agent’s goal should not approve every action it takes. Today, I added separate approvals for write and destructive actions in Agent Workspace. Built with Laravel, the Laravel AI SDK, and Composio. The existing approval lets the agent work toward a goal. Each write or destructive action requires an additional approval before execution. The screenshots show a Google Calendar deletion: 1. I ask the agent to delete a test event. 2. The action pauses for approval. 3. The approval form shows the calendar, event details, and an explicit confirmation that the deletion cannot be undone. 4. After I approve, the agent executes the deletion and returns the result. Even a direct request to delete an event does not bypass this step. That is the distinction I wanted: approving the goal is not blanket permission to change external systems. For those building tool-using agents: do you require approval for every write, or allow some writes within predefined limits? #Laravel #PHP #AIAgents #BuildInPublic

Oct 19

Agent Workspace now executes actions. Built with Laravel, the Laravel AI SDK, and Composio. The flow is now working: 1. Set a goal through a conversation with the agent. 2. The agent asks for approval before acting. 3. Approve, and the agent executes actions toward the goal. 4. Change the goal...Read more

Agent Workspace now executes actions. Built with Laravel, the Laravel AI SDK, and Composio. The flow is now working: 1. Set a goal through a conversation with the agent. 2. The agent asks for approval before acting. 3. Approve, and the agent executes actions toward the goal. 4. Change the goal, and the agent asks for approval again before proceeding. So far, I have tested GitHub and Google Calendar. Only read-only actions. This moves the project beyond defining goals and connecting accounts. The agent can now use those connections to carry out a goal. An important boundary is still unfinished: the system does not yet distinguish read-only operations from write operations. Testing only reads is not the same as enforcing read-only access. Next, I will add separate approvals for write operations, including destructive actions. The current approval flow is working, but that additional control has not been implemented yet. For those building tool-using agents: do you separate approval of the goal from approval of individual write operations? #Laravel #PHP #AIAgents #BuildInPublic