Yoram Kornatzky

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

Oct 14

Agent Workspace now has a working Composio connection. I took two small steps rather than trying to build the entire connector flow at once. First, I merged configurable per-user connectors at the UI level. No Composio yet. A small bite to get out of the reading loop and back to building. Then I ...Read more

Agent Workspace now has a working Composio connection. I took two small steps rather than trying to build the entire connector flow at once. First, I merged configurable per-user connectors at the UI level. No Composio yet. A small bite to get out of the reading loop and back to building. Then I added the connection flow: 1. If a connection is missing, the agent suggests connecting. 2. If a connection exists, it asks which actions to take. 3. The user approves the proposed tool permissions before the agent proceeds. 4. Permissions are isolated per agent. This is progress on the second milestone of my Laravel AI SDK build. But a working connection and an approval prompt are not the same as a completed action. Next is executing a goal end-to-end. For example, listing the inbox or sending a message. That part is not working yet. The useful boundary at this stage is between connecting an account, granting an agent permission, and approving an action. When building tool-using agents, do you prove one action end-to-end first, or build the permission and approval flow before execution? #Laravel #LaravelAI #PHP #AgenticAI #BuildInPublic

Oct 12

Agent Workspace has its first working milestone. Each user can now create multiple agents. Each agent has: 1. Its own goal. 2. Its own chat history. 3. Files uploaded within its conversation. When an agent is created, it asks questions to clarify its goal. The resulting definition becomes its pr...Read more

Agent Workspace has its first working milestone. Each user can now create multiple agents. Each agent has: 1. Its own goal. 2. Its own chat history. 3. Files uploaded within its conversation. When an agent is created, it asks questions to clarify its goal. The resulting definition becomes its prompt. I deliberately kept connectors out of this milestone. The original scope included Composio, tool-level permissions, approvals, shared memory, delegation, and audit trails. Too much at once. The next milestone will add connectors and agent-level permissions. Later, predefined agents will be available through the Agent Library. For now, the basic boundary is working: separate agents, separate conversations, separate files. How do you decide when an agent milestone is small enough to build, but still large enough to teach you something? #Laravel #PHP #AIAgents #BuildInPublic

Oct 9

My next Laravel AI build is taking shape. Working name: Agent Workspace. Each user gets a workspace and can create multiple bots. Each bot is a Laravel AI SDK agent with its own chat and instructions. A bot may also delegate work to sub-agents. The workspace will connect to external services thro...Read more

My next Laravel AI build is taking shape. Working name: Agent Workspace. Each user gets a workspace and can create multiple bots. Each bot is a Laravel AI SDK agent with its own chat and instructions. A bot may also delegate work to sub-agents. The workspace will connect to external services through Composio. Initial connectors: 1. Gmail 2. Google Calendar 3. GitHub 4. Notion There will be no cloud computer or browser. The project is about the Laravel control plane around tool-using agents. Before creating the repository, I need to decide several boundaries: Should connectors belong to the workspace, with permissions granted to individual bots, or should every bot have separate connections? Should permissions be broad read/write access or capabilities for individual tools? Should files and memories belong to the workspace, to individual bots, or support both private and shared scopes? What authority should a sub-agent inherit from the bot that creates it? The goal is to explore multi-tenancy, OAuth connections, capability-based permissions, memory, files, delegation, approvals, queues, audit trails, and testing with the Laravel AI SDK. No repository yet. First, the authority and isolation model. Where would you draw these boundaries? #Laravel #LaravelAI #AgenticAI #PHP #Cursor

Oct 6

Cursor is adding Projects with persistent context, coordinator agents, recurring work, and integrations. This will put pressure on a growing category of separate tools built to provide memory and orchestration around coding agents. The durable products will need to do something Cursor cannot easil...Read more

Cursor is adding Projects with persistent context, coordinator agents, recurring work, and integrations. This will put pressure on a growing category of separate tools built to provide memory and orchestration around coding agents. The durable products will need to do something Cursor cannot easily absorb: work across development environments, hold company-wide context, or govern memory independently of one coding platform. How many tools around AI coding agents are products, and how many are features waiting to be absorbed by the IDE? #Cursor