Skip to content

Consistency Models for Agent State

Problem Statement

Agent systems maintain shared state: conversation history, tool results, decisions, context. Traditional consistency models assume deterministic state updates. But agents are non-deterministic — same input produces different outputs depending on temperature, seed, and context.

What does "consistency" mean for agent state when the agent's own behavior is probabilistic?

Solution Approach: Behavioral Consistency

Define consistency in terms of agent behavior, not state identity:

  • Behavioral consistency: agent behavior (latency, token usage, quality) remains predictable even if state replicas diverge
  • State consistency: conversation history, tool results, execution logs eventually converge across replicas

For Edge+Cloud Hybrid:

Scenario Strategy
Edge agent reading conversation history Eventual consistency (local cache + async sync with cloud)
Two edge agents coordinating Causal consistency (one sends result, other reads in happened-before order)
Agent swarm aggregating results Quorum consistency (wait for majority to report)
Workflow checkpointing Strong consistency at boundaries (all agents agree state is saved)

When to Use

Eventual consistency for: - Conversation history (delay acceptable) - Observability logs (async sync ok) - Metrics (freshness lag tolerable)

Causal consistency for: - Coordinating results between agents - Workflow steps (order matters) - Multi-hop messages

Strong consistency for: - Financial transactions - Critical decisions with approval - Checkpoint boundaries

Trade-offs

Model Pros Cons Edge+Cloud
Eventual Fast; high availability Stale reads; convergence delays Good (local caching)
Causal Respects causality Overhead (vector clocks) Good (agent messages ordered)
Strong Simplest to reason about Slow; fails on partition Poor (can't coordinate across partition)
Quorum Balances availability Slower than eventual Fair (quorum within edge, eventual edge-to-cloud)

Recommendation: Use eventual as default; upgrade to causal when ordering matters; strong only at critical boundaries.

Observability Hooks

Metrics: - Replication lag (target: < 100ms edge; < 1s edge-to-cloud) - Consistency violations (target: 0) - Causal order violations (target: 0)

Queries: - "What state did agent X see at time T?" - "Did agents A and B see consistent data?" - "How long to sync after partition healed?"

Failure Scenarios

State divergence: Edge and cloud have different history. - Recovery: Designate edge as primary; replay from edge; merge non-overlapping updates

Causal order violation: Agent B reads state A hasn't sent yet. - Recovery: Vector clocks detect; reject stale reads

Quorum disagreement: Swarm reports conflicting results. - Recovery: Force re-sync from authoritative quorum

References