Siddhant Khare

im fully convinced that LLMs are not an actual net productivity boost (today) they remove the barrier to get started, but they create increasingly complex software which does not appear to be maintainable so far, in my situations, they appear to slow down long term velocity

Agent tooling is moving fast. The part that has not caught up yet is how we think about permissions across teams. When one engineer uses an agent, permissions are simple. You know what the agent can reach. You trust your own judgment. When three teams use agents across a shared codebase, the question changes. It is no longer "can this agent access this file." It is "which team's agent should have access to which service, under what conditions, and who decides." That is an organizational question, not a technical one. The way I think about it: teams own services. Engineers belong to teams. Agents act on behalf of engineers. These are relationships. And relationships are how permissions should work. "Agent belongs to team:payments. team:payments owns service:checkout. Therefore agent can write to service:checkout." A new engineer's agent inherits the boundaries of their team. A cross-team refactoring agent gets temporary, scoped access that expires when the task ends. An overnight agent gets the narrowest access the task requires. An orchestrator can share read access with sub-agents, but not write. The permissions mirror the org. When someone changes teams, the agent's access changes with them. When a service changes ownership, the permissions follow. No policy files to update manually. No role explosion. This is not a new idea. Google's Zanzibar has modeled permissions as relationships at massive scale for years. The insight is that the same pattern applies naturally to agents, because agents operate within the same org structures humans do. I wrote a chapter on this in the Agentic Engineering Guide. It covers the model, the code, and how it plays out across teams.

strace for AI agents 👀

Agent tooling is moving fast. The part that has not caught up yet is how we think about permissions across teams. When one engineer uses an agent, permissions are simple. You know what the agent can reach. You trust your own judgment. When three teams use agents across a shared codebase, the question changes. It is no longer "can this agent access this file." It is "which team's agent should have access to which service, under what conditions, and who decides." That is an organizational question, not a technical one. The way I think about it: teams own services. Engineers belong to teams. Agents act on behalf of engineers. These are relationships. And relationships are how permissions should work. "Agent belongs to team:payments. team:payments owns service:checkout. Therefore agent can write to service:checkout." A new engineer's agent inherits the boundaries of their team. A cross-team refactoring agent gets temporary, scoped access that expires when the task ends. An overnight agent gets the narrowest access the task requires. An orchestrator can share read access with sub-agents, but not write. The permissions mirror the org. When someone changes teams, the agent's access changes with them. When a service changes ownership, the permissions follow. No policy files to update manually. No role explosion. This is not a new idea. Google's Zanzibar has modeled permissions as relationships at massive scale for years. The insight is that the same pattern applies naturally to agents, because agents operate within the same org structures humans do. I wrote a chapter on this in the Agentic Engineering Guide. It covers the model, the code, and how it plays out across teams.