Elements of Agentic Systems Design
William Chen's framework decomposing agentic systems into ten behavioral elements: Context, Memory, Agency, Reasoning, Coordination, Artifacts, Autonomy, Evaluation, Feedback, and Learning. Positioned as a design space map for builders of agent frameworks, SDKs, and platforms. The core thesis: "The model is a stateless text-to-text function. Everything else is architecture you build around it."
Key Quotes
"The model is a stateless text-to-text function. Everything else is architecture you build around it."
Intelligent-seeming behavior traces to concrete code patterns, not emergent model properties.
Key Themes
The ten elements form a useful vocabulary for analyzing any agent system:
- Context -- what the model sees per call (see Experience Design for Agents on drift)
- Memory -- external storage for selective retrieval (see Chief of Staff's three-layer memory)
- Agency -- translation from text to effects (the tool layer)
- Reasoning -- how calls compose (chaining, looping, branching)
- Coordination -- multi-agent communication (see Cord)
- Artifacts -- shared persistent state (see LLM Wiki's wiki pages)
- Autonomy -- what triggers execution (see Ralph's loop, MimiClaw's heartbeat)
- Evaluation -- measuring success (see LLM Evals, Benchmark Exploitation)
- Feedback -- signals steering current behavior
- Learning -- feedback that persists to change future behavior
The externalization pattern is a recurring insight: Memory externalizes Context, Artifacts externalize Coordination, Learning externalizes Feedback. Each element has a "persisted" version that survives across sessions.
Critical Analysis
Strong: the taxonomy is genuinely useful for comparing agent architectures. You can point at any system in this batch and identify which elements it focuses on and which it ignores. The "behavior-to-code mapping" framing keeps it practical rather than academic. CC BY 4.0 licensing encourages adoption.
Weak: ten elements is arguably too many -- some feel like splits of a natural pair (Feedback/Learning, Context/Memory). The framework is descriptive rather than prescriptive: it tells you what elements exist but not how to choose between design options within each element. And the claim that "intelligent-seeming behavior traces to concrete code patterns, not emergent model properties" is partially wrong -- Emotion concepts and their function in a large language model shows that models develop internal representations that influence behavior in ways not predicted by the architecture.
Most useful as a shared vocabulary when discussing agent systems, less useful as a design guide when building them.
For a pattern catalog organized by constraint rather than taxonomy — the Alexander-esque "navigate from problem to solution" model applied to agent design — see Agentic Design (Pattern Catalog), KORTEXYA's free catalog of 280+ connected agent design patterns.
Sources: Elements of Agentic Systems Design Last updated: 2026-08-01