Several threads come together.

How did we get from language models to agents?

TRY THISOpen a milestone to see what it contributed—and what it did not establish.

SELECTED HISTORICAL ANCHORS2017 — 2025

Architecture

Instructions & tools

Evaluation

12 June 2017 · Paper submission

A different way to process context.

The Transformer paper proposed an attention-based architecture, evaluated on translation.

Why it matters here

It provides architectural context for the processing layers in our diagram.

What it didn’t establish

It was not a demonstration of an autonomous agent, or the invention of attention itself.

Vaswani et al.: Attention Is All You Need

Source-linked historical milestones

What this helps you see

Architecture, instruction-following, tool use and evaluation developed along connected paths. These are selected anchors, not a complete history or an inevitable ladder of intelligence.

Keep the limits in view.

Connections help explain the journey; they do not prove direct historical causation or establish who invented a capability first.

Follow the sources

Vaswani et al.: Attention Is All You Need OpenAI: Better language models and their implications Ouyang et al.: Training language models to follow instructions with human feedback Yao et al.: ReAct: Synergizing Reasoning and Acting in Language Models METR: Task-completion time horizons

Ask about what you’re seeing

Authored learning guide · no live AI generation or web search

NEXT QUESTION

How reliably can an agent complete a task?