## 0010 / 2026-09-07
We are the bottleneck.
Humans are the ones slowing agents down, not the other way around.
After using Astra for just a few hours, you can really feel that this model can do A LOT. It doesn't need hand-holding, it doesn't need careful double-checking, it just knows what to do.
For the longest time, the AI was the bottleneck.
Every model since GPT 3.5 had some major shortcomings. Even as recently as Opus 5, we had a model that was totally unhinged and unpredictable... you couldn't trust the model to do stuff for you, autonomously, without supervision.
This, however, will rapidly change.
We will see more and more people delegate parts of their work and life to agents. Astra can already handle end-to-end tasks reliably. I'm finding myself giving it more and more autonomy, over an increasing range of my tasks.
I believe -- and this is a prediction -- that within 2-3 months we (people at the cutting-edge of AI) will have majority of our tasks done by agents autonomosly. Over 50% of the work you do in a single day, will be done by agents running by themselves. They will start themselves, they will do the task just like you want it done, and they will deliver the work.
You will not be in the loop.
Right now, I'm putting more and more of my time into training Agents. Just like a business owner has to train a new employee, and show him what to do (and what not to do!), I am talking to Astra and having it unblock other agents.
Think about it -- if we want to run thousands of agents productively, which is only a matter of time, we have to have Agents orchestrating & unblocking other agents. You, if you are a Human, cannot possibly have 1,000+ agent threads in your mind at all times. Even after 50+ you begin to go crazy... trust me, I tried.
So in order for us to unlock the next order of magnitude of how many agents exist, we must begin delegating our judgment and decision making to agents. This sounds hard. It sounds like a big ask. However, if you want to remain at the frontier of AI it must be done.
There are two main challenges to this...
(1) The Trigger. When does the agent start? What triggers the creation of a new agent? Or the wake-up of an existing agent?
and (2) The Execution. How do you make the agent do the task exactly as you want it done? How you teach the agent? How do you give him the right context & tools?
These are the two biggest roadblocks in the way, however they will be rapidly solved. I'm working on my own multi-agent system just for this purpose, which I intend on open-sourcing in the coming days.
But now matter what happens, or what the future of work looks like, one thing is for certain...
We, humans, are now the bottleneck.
- written by David Ondrej (manually, no agent involved)