To understand why agentic coding is the proving ground, and what determines the sequence of domains that follow, you have to look beneath the product layer at the training paradigm that made it all possible.
Stage 1 — Pretraining (~2020): Raw pattern learning from massive text corpora. Broad knowledge but no ability to follow instructions or reason reliably.
Stage 2 — Supervised Finetuning (~2022): Learning by imitation—showing the model correct input-output pairs. Limited to what humans can demonstrate.
Stage 3 — RLHF (~2022): Reinforcement Learning from Human Feedback for tone, safety, helpfulness. Bottlenecked by expensive, noisy human preferences.
Stage 4 — RLVR (~2025): Reinforcement Learning from Verifiable Rewards. Instead of “which answer looks better?”, programs and test suites verify: “is this answer objectively correct?”
Why RLVR Was the Breakthrough
Three properties set it apart:
Non-gameable objective rewards. Code either passes the test suite or it doesn’t. This allows much longer optimization runs.
Emergent reasoning strategies. Models spontaneously developed problem-solving approaches—breaking problems into steps, trying multiple approaches, backtracking from dead ends. The model discovers its own strategies rather than imitating human ones.
High capability per dollar. Compute redirected from pretraining into longer RL runs against verifiable rewards became the primary driver of capability progress.
The Verifiability Spectrum: Why Code Fell First
The feasibility of agentic AI in any domain maps directly to where that domain sits on the verifiability spectrum:
Tier 1 — Binary Deterministic Verification (Code, Math): The verifier is a program. The answer is objectively right or wrong. This is where RLVR started and where agents are most mature.
Tier 2 — Structured Verification with Reference Answers (Medicine, Chemistry, Finance, Law): Correct answers exist but are embedded in unstructured formats. Research shows RLVR extends successfully across medicine, chemistry, psychology, economics.
Tier 3 — Partial Verification with Rubric-Based Rewards (Complex Reasoning, Strategy, Design): No single correct answer, but quality can be decomposed into verifiable sub-criteria. The RLVRR framework transforms single-point supervision into “reward chains.”
Tier 4 — Subjective Quality (Creative Work, Brand Voice, Negotiation): No objective verification possible. RLHF remains necessary and agents require the most human oversight.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
