Failure PatternDecision layer
The RADLAB LLC Flat Trust Tier Trap: Why Agencies Fail With Bourdon in Multi-Client Agent Fleets
Symptom: A memory fact written by one client's Cursor agent surfaces inside a second client's Claude session, forcing an emergency audit of what each account could see. Root cause: Bourdon federates memory across Claude, Codex, Cursor, Copilot, Devin, Gemini and Hermes by design, so a single shared store without per-tenant separation lets one client's knowledge hydrate another client's agent.
By InnovaAI ResearchPublished
How do you recognize it?
- •A memory fact written by one client's Cursor agent surfaces inside a second client's Claude session, forcing an emergency audit of what each account could see
- •Every agent in the fleet runs on the same per-agent token with default trust tiers because nobody revisited access control after the initial MCP config
- •Recall feels instant but agents answer with stale or wrong-client context, and the team blames the recognition-first layer instead of the shared graph
- •Onboarding a new client takes days of adapter and MCP config work, yet the agency still cannot show the client a per-tenant isolation boundary
- •The $12 USD per seat monthly line looks trivial until seat count grows across every client agent and nobody has mapped seats to billable retainers
Why does it happen?
- •Bourdon federates memory across Claude, Codex, Cursor, Copilot, Devin, Gemini and Hermes by design, so a single shared store without per-tenant separation lets one client's knowledge hydrate another client's agent
- •Per-agent tokens and trust tiers exist from day one but the medium setup complexity means most agencies finish the MCP config and adapter install and never return to scope access
- •Recognition-first recall returns an immediate match while details hydrate in the background, so a misconfigured graph produces confident but wrong-client answers rather than an obvious miss
- •The Hosted Managed plan at $12 USD per seat per month prices seats, not tenants, so agencies that never modeled client boundaries also never modeled the cost of scaling them
How do you fix it?
- •Audit every per-agent token in the Bourdon runtime and split tokens by client tenant before the next agent session starts
- •Rebuild the knowledge graph so each client gets its own tenant boundary, then re-run a write-then-read test across two different agents to confirm isolation
- •Re-tier trust levels so read-only agents cannot write to the shared store, reserving write access for the agents that own that client's workflow
- •Reconcile seat count on the $12 USD per seat Hosted Managed plan against active client retainers and drop seats tied to retired agent configs
More on RADLAB LLC
- StrategyWhy RADLAB LLC Bourdon Turns Agent Fleets Into Retainer Infrastructure
- ConceptRADLAB LLC Seat Economics Threshold
- Evaluation RuleRADLAB LLC Rule: Self-Host Bourdon Before You Resell It to Any Client
- Decision FrameworkRADLAB LLC: Buy vs Skip (Multi-Agent Memory Federation)
- Implementation BlueprintRADLAB LLC Multi-Agent Memory Federation Rollout (7-10 days)
- Operating ProcedureRADLAB LLC Agent Fleet Memory Federation Rollout (Onboarding)
More for Agent Memory Knowledge Connectors
- Failure PatternsWhy Agencies Fail With Knownbase in Multi-Client Delivery
- Failure PatternsWhy Agencies Fail With OzBrain in Multi-Agent Workflows
- Failure PatternsThe Write-Only Memory Trap: Why Agent Memory Connectors Stall Agency Delivery
- Failure PatternsThe Single-Agent Memory Trap: Why Shared Knowledge Connectors Fail Across Agency Delivery Teams