Unified Agent Dynamics (UAD)¶
Source: Huiwen Han, Unified Agent Dynamics: A Multi-Layer Socio-Technical Framework for Enterprise AI Agents (2026), doi:10.5281/zenodo.19822201.
Enterprise agents are not isolated tools. They are socio-technical actors whose cognitive, computational, and organizational dynamics interact.
The three layers¶
UAD models an enterprise agent across three layers:
| Layer | What it covers | Questions it answers |
|---|---|---|
| Cognitive (LC) | Goals, knowledge, reasoning, memory, context | What does the agent know and intend? |
| Computational (LP) | Models, tools, runtime, control, execution | How does the agent act? |
| Organizational (LO) | Roles, authority, incentives, governance, lifecycle | Who owns it, and under what rules? |
Most agent engineering focuses on LC and LP. UAD’s central claim is that the organizational layer is not an external wrapper; it participates in the agent’s dynamics.
The dynamics equation¶
UAD proposes a canonical form:
dF/dt = G_int(F) + G_env(F, E) + G_ctrl(F, C)
Fis the agent’s state.G_intis internal dynamics (reasoning, memory, planning).G_envis interaction with the environmentE.G_ctrlis the governance/control termC.
The important move is that governance is a term in the dynamics, not an afterthought applied after the agent has already acted.
Three consequences¶
The paper derives three governance-relevant properties:
- A critical governance threshold. Beyond a certain enforcement intensity
μc, stronger control can destabilize behavior (a Hopf bifurcation). More control is not always more stable; control design matters. - Stratification. Heterogeneous enterprise multi-agent systems tend to self-organize into hierarchical layers that can be predicted from the interaction structure.
- A complexity ceiling. Governance-enforced optimal control has a capacity limit; exceeding it is associated with governance failure.
Treat these as theoretical results from a simulation study, not as universal constants. Their practical value is directional: governance has dynamics, and complexity has limits.
Five enterprise agent types¶
UAD distinguishes five types: Agent as Application, Service, Employee, Physical Twin, and Human Twin. The types differ in autonomy, identity, accountability, and failure consequences. See the agent taxonomy for the practical version.
How it maps to AxisVela¶
| UAD layer | AxisVela examples |
|---|---|
| Cognitive | Mnemovela (memory), Gnosivela (grounding), Ontovela (world model), Orchadyn (planning) |
| Computational | PRAXO-CORE (loop and primitives), PRAXOVELA (runtime), Noetivela (inference), Limenora (tools) |
| Organizational | Enterprise Autonomy (composition), Symbivela (human control), Aegivela/NOMIVELA/EIDOVELA (authority), COORDIVELA (work coordination) |
The map is illustrative. A product can contribute to more than one layer.
What to take away¶
- Do not design governance as a post-processing step.
- Model the organizational layer explicitly: roles, authority, lifecycle, and accountability.
- Expect non-linear effects: too little control is unsafe, but poorly designed control can also destabilize.
- Match the control regime to the agent type.