The Five Concerns of Fleet Engineering

June 10, 2026 · View on GitHub

Fleet engineering is organized around five interconnected concerns. Together with a registry, they turn agent populations into operable fleets.

1. Topology

How agents relate structurally.

PatternShapeGood for
HierarchicalManager → workersCommand, delegation, oversight
Peer-to-peerDirect agent-to-agentResilience, emergent coordination
BlackboardShared state / memoryResearch, collaborative reasoning
RouterDispatch to specialistsMulti-domain intake
Market-basedBidding / scoring for tasksDynamic load, cost optimization

Topology choice affects failure modes. Hierarchical fleets fail at the manager; P2P fleets fail at coordination drift.

Fleet engineering question: Who can talk to whom, and who owns the outcome?

2. Choreography

How work flows between agents.

StyleMechanismRisk
Workflow / DAGDefined sequenceBrittle handoffs if schemas are loose
Event-drivenPub-sub, triggersOrdering and duplicate handling
HandoffTyped payload between agentsSchema validation at boundaries
SagaCompensating transactionsLong-running business processes

GitHub's guidance applies at fleet scale: multi-agent systems behave like distributed systems. Every handoff needs typed schemas and boundary validation.

Fleet engineering question: What is the contract at each handoff?

3. Identity & Trust

Who acted, on whose behalf, with what authority.

Minimum fleet identity model:

DimensionOptions
ActorWhich agent instance / version
PrincipalService account (claw) vs end user (assistant)
ScopeTools and data the principal may touch
EvidenceTrace ID, signed action manifest, audit log

LangSmith Fleet implements this as agent identity + credentials (claw vs assistant) + tiered permissions.

Fleet engineering question: Can we prove who authorized this action after the fact?

4. Resource Economics

How scarce resources are allocated across the fleet.

Resources to meter:

  • Tokens / inference spend
  • Tool and API calls
  • Concurrent agent slots
  • Human review bandwidth (inbox queue depth)

Controls:

  • Per-agent and per-team budgets
  • Admission control (reject new runs when over cap)
  • Cost attribution tags (team, agent, loop)
  • Cheap triage before expensive sub-agents

Fleet engineering question: Who pays when a loop runs at 5-minute cadence with two sub-agents?

5. Sovereign Control

How the organization retains ultimate authority.

ControlPurpose
Kill switchEmergency stop for agent / team / fleet
RollbackRevert agent config or action to known-good state
Autonomy tiersF1 supervised → F3 unattended by risk class
Policy-as-codeDenylists, tool allowlists, data boundaries
Drift detectionAgent behavior diverges from registered manifest

Fleet engineering question: Can we stop and recover without guessing?

Registry — The Spine

Without a registry, the five concerns are documentation. With a registry, they are enforceable.

Minimum registry fields per agent:

id: vendor-intake
owner: ops-team
version: 1.2.0
topology: router
identity: claw
permissions: [run]
loops: [daily-triage]
budget_daily_tokens: 500000
autonomy_tier: F1

Template: templates/AGENT-MANIFEST.yaml

How concerns map to patterns

PatternPrimary concerns
Team Agent RegistryRegistry, Identity
Shared Inbox HITLChoreography, Sovereign control
Hierarchical DelegationTopology, Choreography
Agent Clone & ForkRegistry, Permissions
Fleet Budget GuardEconomics
Cross-Agent AuditIdentity, Sovereign control