Community Language Signals Where AI Attention Moves — Cobus Greyling
Why this is in the vault
This is the meta-frame for the 6+ harness-thesis sources already in the vault. Greyling doesn’t just argue that harnesses matter — he explains why the conversation shifted there by tracking the language the community uses. The argument: practitioners sequentially exhaust lower layers and move outward, and the terminology they adopt is the leading indicator of where the center of gravity lands next. This is a Sanity Check article candidate: “The Harness Era.”
Core thesis: a three-layer timeline
Greyling identifies three sequential phases of attention in the AI builder community, each marked by a distinct vocabulary:
-
Weights era (2022) — Community language centered on model internals: parameters, fine-tuning, RLHF, alignment, scaling laws. The center of gravity was inside the model.
-
Context era (2023-2024) — Language shifted to RAG, vector databases, prompt engineering, context windows, retrieval. The center of gravity moved to what surrounds the model call — the information fed in and out.
-
Harness era (now, 2025-2026) — Language has shifted again to agent harnesses, skills, memory management, orchestration, tool calling, protocols. The center of gravity has moved outward to the infrastructure orchestrating the model.
The pattern: each layer gets exhausted (commoditized, well-understood) and practitioners move their energy to the next layer out. The terminology the community adopts is the signal.
Key references
- Cites arxiv paper 2604.08224 — “Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering” (Zhou et al.) — which traces the same weights-to-context-to-harness progression academically.
- Cites Harrison Chase (LangChain CEO) blog post on harness ownership and memory lock-in, which argues harnesses are permanent and tightly coupled to agent memory.
RDCO mapping
- Convergence signal: This is now the 7th+ independent source landing on the harness thesis from a different angle. Greyling’s contribution is the meta-layer — he’s not building harnesses, he’s observing the community’s linguistic migration and naming the pattern.
- Sanity Check candidate: “The Harness Era” — use Greyling’s three-layer timeline as the narrative spine, with the arxiv paper as academic backing and the vault’s practitioner sources (Tan, Chase, Pachaar, Thompson, Klaassen) as evidence.
- Implication for RDCO: The shift is real and durable because it follows the same pattern as previous shifts. Bet on harness-layer tooling and content.
Related
- 2026-04-11-garry-tan-thin-harness-fat-skills
- synthesis-harness-thesis-dissent-2026-04-12
- paper-arxiv-2604-08224-agent-harness-study-2026-04-12
- 2026-04-12-harrison-chase-harness-blog
- 2026-04-10-akshay-pachaar-agent-harness-anatomy
- cross-check-agent-architecture
- 2026-04-26-every-codex-moves-beyond-coding — Every editorializing the same shift (Codex as general work agent, terminal-pilled framing)