Summary

Current AI coding agent frameworks—Orchestrated Agents (like Superpowers) and Role Agents (like Gas Town)—share a fundamental flaw: meta-learning stays with the human, not the agents, who are treated as 'smarter compilers' that execute predetermined plans. Khailo proposes a third model, Learning Agents, embodied in his Abject project, where agents participate in a persistent learning loop using predictions, error signals, and a pattern-language knowledge base that captures both human and machine-acquired intuition. Drawing on six layers of biological and cultural learning systems—from evolution to science—he argues that including agents in the meta-learning loop is the next frontier of agentic software development.

Key Insight

The defining limitation of today's AI coding frameworks is not agent capability but architecture: by reserving meta-learning for humans and treating agents as deterministic executors, we forfeit the compounding gains that come when machines participate in the same prediction-error-revision loops that power every intelligent system from evolution to science.

Spicy Quotes (click to share)

• 9

  A loop without memory is a treadmill. A loop with memory is a ratchet.

• 5

  When the plan itself was wrong, only the person learned.

• 6

  If we treat agents as fancy compilers, we cut them out of that interaction entirely.

• 7

  LLM + harness + memory is an intelligent system. The LLM core is like a processor in a larger system with memory.

• 8

  In my experience, not yet, but the answer won't come from tighter plans but better ratchets.

Tone

analytical, critically optimistic