Prompt Engineering 2022 – 2023 optimizes how you ask Role, few-shot examples, chain-of-thought — turning the ask into a craft. Every technique is the same move underneath: make the output you want the statistically natural continuation of what the model reads.
However well it answers, it cannot see your code, your data, or yesterday's meeting — so putting the facts into the window became its own discipline.
Context Engineering 2023 – 2024 optimizes what the model knows Retrieval, RAG, memory: the right facts in the window at the right moment. The craft is editing, not stuffing — attention sags in the middle of a long window, and every token bills.
Knowing everything, it could still only talk: every conclusion had to be carried back into reality by hand — so the next layer gave the model hands.
Harness Engineering 2024 – 2025 optimizes how the model works Tool calls, permissions, sandboxes, gates: the harness senses and acts so the model never has to leave the text world — and a repo law like AGENTS.md is what those hands obey on every round.
Hands work one round at a time: a miss just stops and waits for a push — so the next layer made the rounds into a loop.
Loop Engineering 2025 – 2026 optimizes how the model keeps going Plan, execute, verify, correct — failure becomes the next round's input. One law is load-bearing: the stop condition must be a gate signal, never the model's own confidence.
One context cannot hold a whole codebase, and one brain runs serially — so the next layer arranged many of them into a graph.
Graph Engineering 2026 – the frontier optimizes how agents cooperate Fan-out for coverage, independent verifiers for trust, pipelines for throughput. Two axioms hold the layer up: parallel value comes from differing perspectives, and a verifier must never be the producer it checks.
every ceiling is the next stage's foundation · click a stage to read why it exists
