November 30
ChatGPT Launches PROMPT ERA
100M users in 2 months. The "prompt engineering" era begins.
Q1 -- Q2
LangChain + LlamaIndex Emerge
RAG becomes the standard pattern for grounding LLMs in external knowledge.
November
GPT-4 Turbo: 128K Context
4x expansion. "Just put it all in context" school of thought forms.
February
Gemini 1.5: 1M Token Context
Paradigm shift. Entire codebases fit in one prompt. RAG vs. long context debate intensifies.
November
Anthropic Releases MCP CONTEXT ERA
Open protocol for LLM-to-tool integration. The "USB for AI" is born.
January
DeepSeek R1: Open-Source Inflection
Matches o1 on reasoning. MIT license. Trained for under $6M. The model is no longer the moat.
Mid-2025
Karpathy: "Context Engineering" > "Prompt Engineering"
The discipline evolves from crafting prompts to assembling entire context pipelines.
July
Manus Publishes Context Engineering Lessons
KV-cache hit rate = key metric. 100:1 input/output ratio. Becomes practitioner reference.
October
Anthropic Launches Agent Skills
Reusable, composable capabilities in YAML + markdown. Skills formalize what power users built informally.
December
A Defining Month HARNESS ERA
Skills open standard (agentskills.io) + MCP joins Linux Foundation + Progressive disclosure adopted + Meta acquires Manus ~$2B
January
ERNIE 5.0 + Kimi K2.5 Agent Swarm
Baidu: 2.4T params. Kimi: spawns 100 sub-agents, 1,500 coordinated tool calls.
February
Heinrich's Skill Graphs Post Goes Viral
8.5K likes, 3.8M views. Skill routing and composition become first-class engineering concerns.
March
Meta-Harness Paper: 6x Performance Gap
Same model, different harness = 6x performance difference. Harness engineering > model selection.
Q1
MCP Reaches 97M+ Monthly Downloads
12x growth in 4 months. The protocol is now infrastructure.
April
This Guide Published
The first unified map connecting every piece of the LLM knowledge engineering ecosystem.