@tjrobertson52

Every week this year, we're getting one step closer to having fully autonomous agents. Not just agents that can do discrete tasks, but agents that are fully autonomous members of your team. Agents working 24 hours a day with no human in the loop.

Earlier this year, OpenAI showed us that we already have the technology to fully close that loop. Now we just need to figure out how to make them reliable, secure, and convenient. And the open source community are doing some great things.

But my guess is that the fully autonomous, reliable, secure, convenient version of the AI agent is gonna be coming from one of the major model makers: OpenAI, Google, or where my money is right now, anthropic.

So I want to talk about two new features that anthropic has added to Claude in the last week that bring us one step closer to these kind of agents. I want to speculate on one other feature that I think is coming soon.

The first feature they've added is the ability to schedule tasks. This, of course, is important for full autonomy because you don't want to have to prompt the agent every time you need a task done. You want it to happen automatically.

We'll, of course, need more advanced triggers for this to be fully autonomous. And eventually we probably need something like open class heartbeat. But task scheduling seems like the logical starting place.

The next feature they've added is remote access. I believe this is currently only available for Claude Code, but you can now communicate with Claude through your phone even when the app is running on your desktop.

If we want these agents to become fully autonomous members of our team, we need to be able to communicate with them where we communicate with our team. This feels like another step towards fully integrating them into our communication channels like email and Slack.

But to get fully autonomous agents, there's another essential feature that none of the model makers have implemented yet. It's part of what makes open class so magical. It's the ability to update its own instructions and knowledge base.

And yes, I know all the major AI models already have memory. However, at least in its current form, memory is just a simple markdown file. It can't possibly contain all the context your agent would need to actually fully autonomous member of your team.

At our agency, we build a brand ambassador for each of our clients, and some of these brand ambassadors now contain hundreds of pages of context. We then provide instructions on how to find the context necessary for a given task. We store all those documents in Google Docs so it's easy to update them and have the brand ambassador update in real time.

The problem is that we're constantly getting new information that we want to add to or update in the brand ambassador, and we still have to manually do that. I just have to believe we're getting close to a world where we could just tell the AI to update its own knowledge base. This is precisely how opencl already works.

And I think that would be a major breakthrough in getting Claude closer to that fully autonomous AI agent.