llm-agent
The LLM-powered autonomous agent — the canonical anatomy from Lilian Weng‘s 2023 survey (2026-06-04-llm-powered-autonomous-agents): an LLM acting as the agent’s “brain” (controller), wrapped by three components.
| Component | What it does | Wiki page |
|---|---|---|
| Planning | Decompose the task into subgoals; reflect on and correct past actions | task-decomposition + self-reflection |
| Memory | Hold the working context (short-term) and recall facts/experience over time (long-term) | agent-memory + maximum-inner-product-search |
| Tool use | Call external APIs for what the weights don’t contain — current info, code execution, proprietary data | tool-use |
The original overview figure puts Agent at the center with four spokes — Tools (Calendar(), Calculator(), CodeInterpreter(), Search(), …), Action, Memory (short-term / long-term), and Planning (Reflection, Self-critics, Chain of thoughts, Subgoal decomposition). That four-box picture became the default mental model for “what an agent is”.
Why this page is the ancestor, not just another concept
This wiki’s center of gravity is 2026-era agent engineering — harnesses, context engineering, verifier loops. Almost all of that vocabulary descends from this 2023 decomposition. The mapping, made explicit so the lineage is visible:
| Weng 2023 (this page) | 2026 wiki descendant |
|---|---|
| LLM brain + perceive/act cycle | agent-loop — the ~20-line ReAct loop (react) |
| Planning → subgoal decomposition | task-decomposition, plan-mode |
| Planning → reflection / self-critics | self-reflection → ralph-wiggum-loop, verifier-loop |
| Short-term memory = in-context learning | context-engineering (working memory) |
| Long-term memory = external vector store | agent-memory, maximum-inner-product-search |
| Tool use (“knowing when & how”) | tool-use → agent-computer-interface, model-context-protocol |
| Chain of Thought as a planning aid | chain-of-thought → reasoning-models |
So the modern stack didn’t replace this frame — it industrialized each box. The harness wraps the loop; context engineering manages the memory budget; ACI/MCP formalize tool use; verifier loops operationalize self-reflection.
What 2026 sources changed
- Memory. Weng’s default was “long-term memory = vector store + RAG”. The 2026 agent-memory sources argue markdown + simple retrieval beats a vector DB at small/medium scale, and split memory four ways (working / procedural / episodic / semantic). Notably, Weng’s human taxonomy already contained episodic + semantic + procedural — the 2026 frame essentially re-applied the human taxonomy to file-backed engineering instead of to a vector store.
- Planning. Reflection here is prompt-/SFT-level (Reflexion, Chain of Hindsight). The 2026 descendants externalize it: agents review each other and loop until verifiers pass (ralph-wiggum-loop).
- Reliability. Weng’s third “challenge” — the natural-language interface is unreliable, so most agent code is output parsing — got addressed by structured tool calling and prompt-cache-shaped harnesses (prompt-caching), not by the problem disappearing.
Weng’s own 2026 revision — agent anatomy → harness anatomy
The author herself updated the frame three years on. 2026-07-07-harness-engineering-self-improvement (Jul 2026) defines the harness as the successor anatomy: everything the 2023 three-component frame had (planning, memory, tool use around an LLM brain), plus workflow design, evaluation, permission controls, and persistent state management — “closer to runtime and OS design than prompt templates”. Two things changed between the frames:
- The unit of engineering moved from components to code. The 2023 anatomy names capabilities; the 2026 harness is a program — “code that programs how prompts, tool calls, subagents, control flow, memory, and workflow logic work together”. That makes the whole anatomy an optimization target (recursive-self-improvement, meta-harness), which the 2023 frame never contemplated.
- Evaluation and permissions joined the anatomy. The 2023 survey treated evaluation as an external research concern; the 2026 frame builds the evaluator (and its placement outside the self-improvement loop) into the system definition itself.
So this page’s lineage table gains a final row: the 2023 anatomy’s descendant is not just each industrialized box — it’s the harness as a whole.
The proof-of-concept era it captured
Written at the AutoGPT / BabyAGI / GPT-Engineer peak. Weng’s verdict on those demos — “a lot of code in AutoGPT is about format parsing” — was the early signal that the harness around the model, not the model alone, was where the engineering would go (harness). Those POCs aren’t their own pages yet; they’re discussed on 2026-06-04-llm-powered-autonomous-agents and tool-use.
Cross-link
- 2026-06-04-llm-powered-autonomous-agents — the source.
- agent-loop — the runtime instantiation of “LLM brain + perceive/act”.
- react — the reasoning + acting primitive the loop runs.
- 2026-04-27-the-second-half-of-ai — the same author-adjacent lineage two years on: the recipe generalized; evaluation became the bottleneck.
- 2026-07-07-harness-engineering-self-improvement — Weng’s own successor frame: agent anatomy → harness anatomy, and the anatomy itself becomes the optimization target (recursive-self-improvement).