Skip to content

Phase 3 — Autonomous Agents

alex-krentsel's four-phase periodization of LLM-system evolution, delivered as the framing device for his 2026 openclaw deep-dive. The version that's useful for the wiki because it draws a sharp line between phase 2 scoped agents (LangChain / AutoGen / CrewAI) and phase 3 autonomous agents (claude-code, openclaw) — and names the discriminator precisely.

The four phases

  • Phase 0 — Next-token predictors. BERT, GPT-1/2/3, LaMDA. Single-token output from a single forward pass. End of Krentsel's undergrad (~2019).
  • Phase 1 — Fine-tuned assistants. ChatGPT / Claude / Gemini. Transformer looped on itself for autoregressive generation, then fine-tuned on human-assistant dialogues to bias chat behavior. ~2021–2022.
  • Phase 2 — Scoped agents with static orchestration. LangChain, AutoGen, CrewAI, Google's early AI Overviews. Tools attached, but control flow is human-authored: "first this agent goes, then this agent goes." What Krentsel calls "static wrappers around a call to an LLM."
  • Phase 3 — Autonomous agents with dynamic tool discovery + orchestration. claude-code, openclaw. Same core LLM + tools, but orchestration and tool selection are made by the model at runtime. OpenClaw goes further: it can modify itself and learn.

Why the Phase 2 / Phase 3 cut matters

The difference isn't capability — phase-2 systems could do many of the same things, given enough human engineering per workflow. The difference is who authors the control flow:

  • Phase 2: human author writes the graph of LLM calls; LLM is a subroutine.
  • Phase 3: LLM authors the graph of LLM calls; human sets the goal.

This is why phase-3 systems can "do things" rather than "do email" or "do calendar" — the generality is a property of where control lives. Krentsel derives OpenClaw's two design goals (closed control loop + flexibility/extensibility) directly from this.

Pairs with

  • agentic-loopiness — each phase wraps a new loop over the previous layer; phase 3 is the matryoshka's outer shell (for now).
  • design-over-implementation — at phase 3, the model handles implementation; humans are left with architecture.
  • harness-engineering — the harness is the phase-3 artifact; in phase 2 it was the whole app.
  • control-flow-vs-prompt-flow — same cut from chris-shayan's banking frame.
  • Krentsel's open question: what is phase 4? His candidate — malleable architecture, systems that self-edit their harness design, not just their skills/tools.

Cross-references