AI
AIEWF Daily Dispatch: Autoresearch and the tension between AI and human agency
At the AI Engineer World's Fair, speakers debated the limits of automation, pushing back against the concept of fully automated "software factories." While
Key takeaways
- Autoresearch allows developers to build loops in which agents help maintain the system itself.
- The outer loop of engineering and agency must remain human.
- AI agents should handle the first 80% of design work, leaving the last 20% for human taste.
- Generative models default to the aesthetics of their modeling teams without creative human input.
- Fully generating entire brand websites with AI is risky because they can stray outside guidelines.
At the AI Engineer World's Fair, speakers debated the limits of automation, pushing back against the concept of fully automated "software factories." While Introspection co-founder Roland Gavrilescu highlighted "autoresearch" loops where agents maintain systems, former Google engineering leader Addy Osmani argued that humans must remain in the outer engineering loop. Other speakers, including Notion's Geoffrey Litt and Impeccable's Paul Bakaus, emphasized that humans must understand their code and retain creative control, warning against delegating complete agency to AI models.
In their words
“There is no auto, and there will be no auto.”
“The models are grown, not developed.”
By the numbers
- 80%
- Laborious design work handled by agents in Impeccable
- 20%
- Final design work reserved for human taste and authorship
- 35.5K
- Views on Geoffrey Litt's tweet criticizing software factories
How it unfolded
- Software factory framing dominates the conference discussions
- Autoresearch and human agency take center stage at AIEWF
Why this matters
-
Those who delegate understanding of code to agents will be replaced by them.
Turn stories like this into views
Ravenclip finds the AI news, makes the video, and posts it before attention moves on.
Common questions
- What happened with AI Engineer World's Fair?
- Autoresearch allows developers to build loops in which agents help maintain the system itself.
- Where can I read the original report?
- Read the full report at latent_space.
