agent persistence & embodiment
on this page
quick read
definitions vary in the kind of “body” and lifespan they assume for agents. this dimension groups them into three descriptive modes:
- ephemeral executors: short-lived scripts or session agents that dissolve after a task.
- persistent digital systems: software with memory, identity, or always-on operation.
- embodied / thermodynamic entities: agents tied to physical bodies or biological processes.
Persistence and embodiment in agent definitions
Rendering diagram...
Representative definitions organized by lifespan and embodiment.
comparison table
| profile | hallmark traits | sample definitions | diagnostic questions |
|---|---|---|---|
| ephemeral executor | in-memory or per-task existence; limited state retention | ChatGPT Agent release (2025), LangGraph workflows (2025), tool-enabled LLM sessions | Does the agent vanish after the conversation or task completes? |
| persistent digital system | maintains identity, memory, or continuous presence | Pattie Maes software agents (1994), Franklin & Graesser autonomous agent (1996), Claude tool use (2025) | Does the agent store context, run continuously, or resume tasks later? |
| embodied / thermodynamic | relies on physical body, sensors, or metabolic work cycles | Russell & Norvig robotic agents (1995), Kauffman autonomous agents (2000), Holland complex adaptive systems (1995) | Does agency depend on a body, energy budget, or environment embodiment? |
notable examples
ephemeral executors
- ChatGPT agent release notes (2025, AI): describes “modes” that complete complex tasks online during a session, then hand back control; the agent is spun up per assignment.
- LangGraph workflows vs. agents (2025, AI): workflows are deterministic graphs triggered per run, while “agents” are still session-scoped orchestrations without guaranteed persistence.
- OpenAI Operator (2025, AI): preview browser agent that performs tasks, then yields; state persists only within the job window.
persistent digital systems
- Pattie Maes software agents (1994, HCI): envisioned always-on assistants that learn over time and continuously filter information for users.
- Franklin & Graesser autonomous agent (1996, AI): requires systems that “persist over time” pursuing agendas, explicitly differentiating them from transient programs.
- Anthropic Claude tool use (2025, AI): iterative tool calls with memory across steps; the agent maintains context to plan subsequent actions within a long-running session or workflow.
embodied / thermodynamic entities
- Russell & Norvig AI agent varieties (1995, AI): include robotic agents with sensors and actuators, highlighting physical embodiment.
- John Holland hidden order (1995, complex systems): agents are adaptive, rule-following entities situated in an environment, often modeled with tangible states or resources.
- Stuart Kauffman autonomous agents (2000, complex systems): defines agents as systems capable of reproduction and thermodynamic work cycles—explicitly physical.
self-critique
- session vs. persistence continuum: modern LLM products sometimes serialize memory; classification may change rapidly as persistence features roll out.
- embodiment ambiguity: software can act through APIs; whether that counts as “embodied” depends on how strictly we require physical bodies.
- mixed realities: agent-based simulations (Epstein & Axtell) and game agents straddle digital persistence and embodied metaphors, complicating buckets.
questions for you
- should we carve out an explicit “cyber-physical hybrid” class for agents with digital brains but robotic bodies?
- how might legal accountability shift when an agent persists across sessions versus being spun up per task?
- which definitions in the dataset emphasize memory or lifecycle management strongly enough to warrant subcategories?
- do thermodynamic notions of agency (Kauffman) belong in the same taxonomy as LLM orchestrators, or should they sit in a separate ontology?