Files
WebinarNotes/wiki/timelines/ai-agent-evolution.md

3.8 KiB
Raw Blame History

Timeline: AI Agent Evolution

#timeline

Chronological reconstruction of how AI agents evolved, combining konstantin's tooling/agent timeline with theo-browne's model eras. Dates are as stated by the speakers (approximate; some are framing rather than precise release dates). Status: tentative where noted.

Agent architecture & tooling (Konstantin)

Period Milestone Significance
Early 2023 LLM / ChatGPT (pre-chat): ask → answer No memory, no tools, single-turn
Summer 2023 ReAct agents — LLM can answer or call a function First "agent loop"; tools give the LLM feedback. See agentic-loops
20232024 Call chains, roles, first agent SDKs, vector DBs, RAG, JSON outputs Agents use external knowledge
2024 Chains → graphs (LangGraph); multi-agent (planner/critic); AutoGen, CrewAI; heavy scaffolding Peak technical complexity
Late 2025 Simplification — a simple agent with ~10 file ops solves hard tasks; era of universal agents = harness Complexity drops, capability rises
2026 (now) Harnesses wrapped in outer loops (Ralph loop), run for days; AI factory (no human-in-the-loop) human-in-the-middle → man-on-the-middle → no human
2027 (forecast) Dominance of skills-as-memory + agentic-loops

Tooling generations (see evolution-of-agent-tooling): Tools (20222024) → MCP (late 20242025) → Skills (2025→).

Human-role ladder (see context-as-scarce-resource): prompt-engineer → context-engineer → harness-builder → loop-engineer.

Model capability eras (Theo Browne)

Speaker's framing; model-name→ID mapping is unverified/tentative (see theo-browne).

Era Representative model (speaker's term) Unlocked
Tool-call era "Sonnet 3.5" First reliable tool calls inside a real codebase
Long-running-task era "Opus 4.5" Multi-hour tasks; self-run/test (his "AI psychosis" began NovDec 2024)
Orchestration era "Mythos / Fable" Understands itself; spawns sub-models; verifies afterward — a prompt replaces a software factory

The two timelines rhyme: Konstantin's "universal harness → loops → AI factory" is the system view of Theo's "orchestration era" model capability. See think-wider-not-bigger.

Uncertainty

  • Konstantin: "everything changes every 6 months — nothing is fundamental yet." Treat 2027 as a forecast.
  • Theo's model-era names are rhetorical; do not treat as official release labels.