Goodbye Slop; Welcome Determinism
summary
Argues that replacing mega‑prompts with explicit state machines removes agent “slop” by separating deterministic rules from generative work and makes behavior inspectable. Implements an email‑drafting agent with rules like must draft before send and never invent a recipient, models it in XState, sends to a simulated outbox in a live demo, logs structured state machine traces, evaluates states and edges, and has an agent propose a V2 workflow that runs faster and reduces unnecessary review loops.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 1:19State machines and deterministic behavior
- 2:11From infant loops to graphs
- 3:00Graphs in agent engineering
- 3:45Where agent slop comes from
- 4:44Defining determinism
- 5:17The email-drafting example
- 6:12Control flow inside a mega-prompt
- 8:07An explicit state machine
- 9:05Live demo: drafting and sending an email
- 10:35Modeling replaces confusion
- 11:42Inspectable structure and observability
- 12:21Evaluating states, outcomes, and edges
- 13:23Asking an agent to improve the machine
- 14:11Version two of the workflow
- 15:33The minimal execution loop
- 16:24Create, observe, improve, compare
- 17:18Related work
- 18:00From guardrails to self-improvement
- 18:45Espresso-machine demo
- 20:20Closing
related talks
42mMaking state management intelligent
19mDon't Build Slop (4 Levels of AI Agent Maturity)
32mDAVID KHOURSHID Making state management intelligent
34mMaking state management intelligent
38mInto the Visual World of State Machines
25mMastering UIs with Finite State Machines
37mKnowing your state machines
28mState machines for frontend state management