Nathaniel Whittemore
👤 SpeakerAppearances Over Time
Podcast Appearances
They wrote, We've been experimenting with running coding agents autonomously for weeks.
Our goal is to understand how far we can push the frontier of agentic coding for projects that typically take human teams months to complete.
And indeed, if you want to take a step back and just try to understand psychologically where the vanguard of AI and agentic coders are right now, it is really all about pushing the boundaries on autonomy.
Breaking out, in other words, of being the bottleneck where without your consistent prompting, the AI isn't doing anything.
The leading agentic coders are in the midst of trying to build systems that work all the time with extremely minimal input from them.
They want nothing less than armies of agents that work while they sleep.
And that army idea is operative.
In that same cursor blog, they write, Today's agents work well for focused tasks but are slow for complex projects.
The natural next step is to run multiple agents in parallel, but figuring out how to coordinate them is challenging.
Initially, Cursor gave their coding agents equal status, and as they put it, let themselves coordinate through a shared file.
Each agent would check what others were doing, claim a task, and update its status.
Ultimately, however, this failed.
The locking mechanism they implemented to prevent two agents from grabbing the same task ended up becoming a bottleneck.
As they put it, 20 agents would slow down to the effective throughput of two or three with most time spent waiting.
They tried a second strategy, where agents could read state freely, but writes would fail if the state had changed since they last read it.
In other words, they couldn't make different updates to the same code at the same time in an attempt to avoid conflicts.
However, Cursor wrote this didn't work either.
Quote, As they put it, with no hierarchy, agents became risk-averse.
They avoided difficult tasks and made small, safe changes instead.
No agent took responsibility for hard problems or end-to-end implementation.