SKILL.md — FIPA Agent Management: Multi-Agent Infrastructure Patterns

metadata:
  name: fipa-agent-management
  version: XC00023H
  description: >
    Canonical patterns for naming, registering, lifecycling, and managing
    autonomous agents in multi-agent systems. Derived from the FIPA Agent
    Management Specification — the normative standard for interoperable
    agent infrastructure.
  activation_triggers:
    - designing multi-agent orchestration systems
    - implementing agent discovery or capability routing
    - building skill registries or service directories
    - handling agent failures or retry logic in distributed systems
    - designing agent identity, addressing, or transport layers
    - modeling agent state or lifecycle transitions
    - building federated or hierarchical agent platforms

Decision Points

Core Architecture Choice

IF building single-organization agent platform (< 50 agents):
  ├─ Use single AMS + single DF
  ├─ Direct resolution of AIDs to transport addresses
  └─ Flat capability search across all agents

ELSE IF building multi-organization/federated system:
  ├─ Use federated DF architecture
  ├─ Each organization runs own AMS/DF pair
  ├─ DFs register with parent/peer DFs
  └─ Search propagates with max-depth bounds

Agent Discovery Strategy

IF agent needs specific capability:
  ├─ Query local DF first with service description
  ├─ IF no results AND federated topology exists:
  │   └─ Propagate search to federated DFs (depth ≤ 3)
  └─ ELSE IF no results in flat topology:
      └─ Return "capability not available" (don't retry)

IF agent joins platform:
  ├─ Register with AMS first (establish identity + lifecycle state)
  ├─ THEN register capabilities with DF
  └─ Registration order matters: existence before capabilities

See references/agent-discovery-without-omniscience.md for why a single central registry eventually fails under real-world dynamic conditions.

Failure Recovery Logic

IF receive typed FIPA exception:
  ├─ unsupported/unrecognised → route to different agent type
  ├─ missing/malformed → fix request parameters, retry once
  ├─ unauthorised → escalate to platform admin
  ├─ not-registered → agent reference stale, re-discover via DF
  ├─ already-registered → idempotent operation, continue
  └─ internal-error → exponential backoff retry (max 3 attempts)

ELSE IF receive non-FIPA error or timeout:
  ├─ Mark agent as Unknown in AMS
  ├─ Trigger platform-level health check
  └─ Do not retry until health restored

Lifecycle State Transitions

IF need to send request to agent:
  ├─ Query AMS for current lifecycle state
  ├─ IF state = Active: send immediately
  ├─ IF state = Waiting: send with acknowledgment protocol
  ├─ IF state = Suspended: queue or route to alternative
  └─ IF state = Unknown/Transit: wait for state resolution

IF agent reports failure/becomes unresponsive:
  ├─ AMS transitions agent to Unknown state
  ├─ DF keeps capability registrations (for recovery)
  ├─ Platform stops routing new requests to agent
  └─ Initiate recovery protocol (restart/reregister)

See references/lifecycle-states-as-coordination-contracts.md for when each unavailability state demands a different orchestrator response.

Failure Modes

🚫 Endpoint Conflation

Symptoms: Direct calls to http://agent.host:8080, hard-coded URLs in routing logic, broken references when agents move
Detection Rule: If you store transport addresses as permanent identifiers, you have this anti-pattern
Fix: Store AIDs only; resolve transport addresses via AMS at invocation time — see references/fipa-identity-vs-location-separation.md for the formal AID model

🚫 Registry Collapse

Symptoms: Single table/store tracking both agent existence and capabilities, lifecycle changes break service discovery
Detection Rule: If updating agent state requires touching capability records, you've collapsed the registries
Fix: Separate AMS (white pages) from DF (yellow pages); independent update operations — see references/fipa-two-registry-architecture.md for the canonical separation argument

🚫 Blind Retry Loops

Symptoms: Generic error handling that retries all failures identically, no differentiation between temporary vs permanent failures
Detection Rule: If your retry logic doesn't branch on failure type, you're in blind retry
Fix: Parse FIPA exception types; different recovery strategies per exception class — see references/fipa-exception-taxonomy-as-reasoning-tool.md for the complete taxonomy

🚫 State-Blind Invocation

Symptoms: Sending requests without checking agent lifecycle state, timeouts on suspended agents
Detection Rule: If you invoke agents without AMS state check, you're state-blind
Fix: Query lifecycle state before invocation; handle each state appropriately

🚫 Discovery Bottleneck

Symptoms: Single central DF handling all capability queries, search latency increases with agent count
Detection Rule: If all capability searches hit the same registry, you have a bottleneck
Fix: Implement federated DF architecture with bounded search propagation

Worked Examples

Example 1: Federated Discovery Across Organizations

Scenario: Agent in Organization A needs text-translation capability, which exists in Organization B's platform.

Step-by-step execution:

1. Agent queries local DF: QUERY-REF(service-type=translation, input-lang=en, output-lang=es)
2. Local DF returns empty result set
3. Local DF has federated peer registered → propagate search with max-depth=2
4. Organization B's DF receives federated query, matches against registered translator agent
5. Result propagates back: AID=translator@orgb.com, service-description=...
6. Agent A resolves AID via Organization B's AMS → gets transport address
7. Agent A sends translation request with REQUEST performative

Novice mistakes: Querying Organization B directly, storing B's transport address permanently, not using AID resolution
Expert insights: Federated search is bounded (prevents query storms), AID resolution happens at invocation time (handles agent migration), performative choice signals intent clearly

Example 2: Agent Failure and Re-registration

Scenario: Active translation agent crashes mid-task, client needs to detect failure and recover.

Step-by-step execution:

1. Client sends REQUEST to translator agent → receives timeout (no response)
2. Client queries AMS for agent lifecycle state → returns "Active" (stale)
3. Platform health monitor detects agent unresponsive → AMS transitions to "Unknown"
4. Client retries request → receives not-registered exception from message transport
5. Client re-queries DF for translation capability → finds replacement agent
6. Original agent restarts, re-registers with AMS (Initiated → Active transition)
7. Original agent re-registers capabilities with DF
8. Subsequent capability queries return both translator agents

Novice mistakes: Infinite retry without state checking, assuming agent death means capability unavailable
Expert insights: AMS state change is separate from DF re-registration, multiple agents can offer same capability, exception type determines recovery strategy

Example 3: Multi-Protocol AID Update

Scenario: Agent platform adds MQTT transport alongside existing HTTP, agent needs to be reachable via both protocols.

Step-by-step execution:

1. Agent updates AMS registration: AID=agent@platform.com, addresses=[http://host:8080/agent, mqtt://broker:1883/agent]
2. AMS validates both transport addresses are reachable
3. Clients continue using same AID for agent reference
4. When client resolves AID → AMS returns both addresses, client chooses transport based on capability
5. HTTP clients continue working unchanged
6. New MQTT clients can connect using same agent identity

Novice mistakes: Creating separate agent identities per protocol, forcing all clients to update references
Expert insights: Identity stability across transport changes, AID supports multiple addresses natively, transport choice happens at resolution time

Quality Gates

• Agent identities are symbolic AIDs, not transport addresses
• AMS (white pages) and DF (yellow pages) are separate registration systems
• All agent-to-agent communication checks lifecycle state before invocation
• Error responses use FIPA exception taxonomy, not generic success/failure flags
• Capability searches support both local and federated discovery modes
• Agent registration follows correct sequence: AMS first, then DF
• Federation topology has explicit max-depth bounds on search propagation
• Lifecycle state transitions respect authority rules (AMS vs agent-initiated)
• Message performatives match communication intent (REQUEST vs INFORM vs QUERY)
• Recovery protocols distinguish between agent failure and capability unavailability

Bundled Assets

• Diagrams — diagrams/INDEX.md: lifecycle state machine, discovery/registration sequence, and agent addressing decision tree
• References — references/INDEX.md: deep-dives on AMS/DF separation, AID identity model, exception taxonomy, federated discovery, capability search, lifecycle contracts, registry authorization, message performatives, and platform interoperability

Not-For Boundaries

Do NOT use this skill for:

• Individual agent reasoning or decision-making → Use [agent-cognition-patterns] instead
• Task decomposition within single agent → Use [task-planning-protocols] instead
• Agent learning or adaptation strategies → Use [agent-learning-frameworks] instead
• User interface design for agent interactions → Use [human-agent-interaction] instead
• Business logic or domain-specific agent behaviors → Use domain-specific skills instead

This skill covers infrastructure-level coordination only: how agents find each other, communicate, and manage shared platform resources. For everything that happens inside an individual agent, delegate to other skills.