← back
Don't Build Slop (4 Levels of AI Agent Maturity) - Ara Khan, Cline
Takeaway
Advance agent infrastructure according to product needs while keeping execution understandable, minimal, and easy to test.
Summary
- Khan proposes four maturity stages: framework-based prototyping, a custom agent implementation, Kanban-oriented workflows, and cloud deployment.
- Frameworks are presented as a quick way to validate whether an agent solves a useful problem.
- Production agents should be understood as state machines with explicit actions, transitions, and completion conditions.
- The talk recommends pruning accumulated prompts and logic as models improve and exposing agents through a CLI that supports repeated building and testing.
clineagent-designagent-maturity
Original description
The prompt for GPT-5.3 is one-third the size of the one written for GPT-5. Frontier models are so capable that longer system prompts cause sensory overload and degrade performance. The rule Ara Khan keeps returning to: every single thing you add to an agent risks making it worse. The talk breaks agent-building into four levels, from framework prototyping to cloud-native fleets, with five concrete rules for writing your own agent code in between. The form factor argument lands on Kanban boards as the right UX for managing parallel inference-bound agents. He made that prediction publicly on March 26. Claude Code shipped the same thing ten hours before this talk. Speaker info: https://x.com/arafatkatze / arafatkatze https://github.com/arafatkatze Timestamps: 0:00 Introduction: Addressing the "mass psychosis" of agent-building 2:13 The Four Levels of Agent Maturity overview 3:00 Level 1: Using existing agent frameworks 4:13 Level 2: Building your own agents (with five rules) 4:36 Rule 1: View every agent as a state machine 5:55 Rule 2: Keep it simple (Avoid unnecessary system prompt bloat) 7:20 Rule 3: Integrate with a CLI for a pseudo-RL pipeline 8:43 Rule 4: Don't build "slop" (Focus on thoughtful architecture) 9:43 Rule 5: Master frontier model APIs (Reasoning traces) 11:03 The UX form factor: Why Kanban boards are best for agents 13:12 Shipping to the cloud for scalability 17:15 Q&A session: Planning and state transitions in Kanban