Skip to content

Multi-agent dynamics

A multi-agent system is not “one agent, repeated.” Adding agents changes the failure modes: coordination cost, shared state, conflicting authority, emergent behavior, and organizational structure all become first-class problems.

What changes at scale

Single agent Multi-agent system
One context and one plan Shared state and divergent context
One authority boundary Delegation chains and conflicting permissions
One failure domain Partial failure and recovery across agents
One trace Correlated traces and causal attribution
One lifecycle Roles, handoffs, leases, and retirement

UAD’s organizational layer is the relevant lens: agents form roles, authority relations, and interaction structures. The paper’s stratification result suggests that heterogeneous systems tend to organize into hierarchical layers predictable from the interaction matrix. Treat that as a design signal, not a law of nature.

Coordination dimensions

Harmovela defines seven dimensions that a multi-agent coordination layer must address:

  1. Event — what happened.
  2. Task — what is executing and in which state.
  3. State — what is true now, and how it changes.
  4. Context / memory — what informs decisions.
  5. Delegation — who does what, with which authority.
  6. Recovery — what happens when something fails.
  7. Governance — who may do what, and what evidence remains.

Most multi-agent failures are a missing answer to one of these dimensions, not a model-quality problem.

Common failure modes

  • Duplicate work. Two agents claim the same task because there is no lease or ownership model.
  • State divergence. Agents hold incompatible views of the world and act on both.
  • Runaway delegation. Authority expands as it passes between agents, or a loop of agents delegates to each other.
  • Coordination overhead. The system spends more effort coordinating than doing.
  • Emergent escalation. Local rules produce a global behavior nobody designed.
  • Attribution loss. When something goes wrong, no one can reconstruct which agent did what under which authority.
  • Complexity ceiling. The governance load grows faster than the system’s capacity to enforce it.

Design rules

  1. Name roles and responsibilities. Do not rely on emergent division of labor.
  2. Use leases and claims. Ownership should expire and be observable.
  3. Share state through a protocol, not side channels. Harmovela or an equivalent coordination layer.
  4. Bound delegation. Authority should attenuate as it moves; see AEGIVELA.
  5. Make handoffs explicit. A handoff is a state transition with an owner and evidence.
  6. Design for partial failure. Recovery is a coordination dimension, not an afterthought.
  7. Watch the governance threshold. More control can destabilize if it is poorly designed; measure the effect.
  8. Keep an attribution trail. Correlated traces and identity are mandatory.

How it maps to AxisVela

  • HARMOVELA — protocol for the seven coordination dimensions.
  • COORDIVELA — coordination of development work: leases, dependencies, conflicts, interventions.
  • RHEOVELA — durable process execution and recovery.
  • SYMBIVELA — human oversight of multi-agent work.
  • NOMIVELA and AEGIVELA — identity and bounded authority across agents.

What to take away

  • Multi-agent design is organizational design.
  • Coordination must be explicit, observable, and recoverable.
  • Delegation without attenuation is a security bug.
  • Measure coordination cost and governance load, not just task throughput.