SKILL.md — Agha Actor Model: Concurrent Agent System Design
Decision Points
Choosing Actor Creation Strategy Based on Task Type
IF task has sequential dependencies:
├─ Use customer pattern: create child with reply address
├─ Pass customer address to child as parameter
└─ Child sends result directly to customer (not parent)
IF task requires long computation:
├─ Create insensitive actor for computation
├─ Forward incoming messages to buffer actor
└─ Resume from buffer when computation completes
IF task needs dynamic resource allocation:
├─ Create resource manager actors on demand
├─ Pass capabilities (addresses) as message data
└─ No central registry - addresses flow through system
IF task has failure isolation requirements:
├─ Spawn supervised child actors for risky operations
├─ Supervisor detects failure via missing replies
└─ Replace failed actors without affecting others
IF composing existing agent systems:
├─ Verify interface preserves causal structure
├─ Test behavior under composition (not just isolation)
└─ Use message protocols as boundaries (not shared state)
Message Sending vs State Replacement Decision Tree
IF coordination needed with other agents:
├─ SEND messages (don't modify local state first)
├─ Include reply address if response expected
└─ Never assume message ordering
IF local computation needed:
├─ SPECIFY replacement behavior
├─ Encapsulate new state (don't expose internals)
└─ Ensure one-message-at-a-time processing
IF dynamic scaling needed:
├─ CREATE new actors with specific behaviors
├─ Pass necessary addresses to new actors
└─ No shared initialization state
Failure Modes
1. Central Orchestrator Anti-Pattern
Detection: If you see one actor routing all messages or holding all system state Symptoms: Single point of failure, bottleneck under load, infinite regression problem Fix: Decompose into community of actors, each knowing only local context, use capability routing
2. Synchronous Blocking Fallacy
Detection: If actors wait/block for responses instead of specifying replacement behavior Symptoms: Deadlock under load, hidden timing assumptions, reduced concurrency Fix: Model as request-reply message pairs, use customer pattern for dependencies, apply insensitive actor pattern
3. Shared State Contamination
Detection: If multiple actors read/write same data structure (even with locks) Symptoms: Race conditions, sequential bottlenecks, hidden global state Fix: Encapsulate state in single actor, use message passing for coordination, mutual exclusion is free
4. Output-Only Verification
Detection: If testing only compares final outputs without checking interaction patterns Symptoms: Brock-Ackerman anomaly - identical outputs but different composition behavior Fix: Verify causal structure preservation, test behavior under composition, use observation equivalence
5. Static Topology Assumption
Detection: If communication graph is fixed at startup with no runtime reconfiguration Symptoms: Cannot handle open systems, no dynamic resource management, brittle under change Fix: Treat addresses as first-class data, implement capability routing, support runtime topology changes
Worked Examples
Example 1: Task Decomposition with Customer Pattern
Scenario: Agent needs to process a complex request requiring sequential subtasks A → B → C, but must remain responsive to other messages.
Novice Approach:
receive request →
block while calling subtask A
block while calling subtask B
block while calling subtask C
send final result
Expert Application of Actor Model:
receive request →
create customer_BC actor with addresses for B, C, final recipient
send subtask A request to A_processor with customer_BC as reply address
specify replacement behavior: ready for next request
customer_BC receives A result →
send result to B_processor with customer_C as reply address
customer_C receives B result →
send result to C_processor with final_recipient as reply address
Key Decisions Made:
- Used customer pattern to avoid blocking
- Each step creates next customer in chain
- Original agent remains responsive throughout
- Failure in any subtask only affects that chain
Example 2: Failure Recovery with Supervision
Scenario: System needs to handle agent failures without cascading to whole system.
Novice Approach: Try-catch around agent calls, restart everything on failure.
Expert Application:
supervisor creates worker_actor →
sends task to worker with reply timeout
specifies replacement: "waiting_for_reply"
IF reply received within timeout →
forward result to client
specify replacement: "ready"
IF timeout expires →
create new worker_actor (old one failed)
resend task to new worker
specify replacement: "waiting_for_reply"
Trade-offs Navigated:
- Supervision is separate concern from business logic
- Failed actors are isolated and replaced, not repaired
- Timeout detection vs guaranteed delivery balance
- No shared state between supervisor and workers
Quality Gates
- No actor blocks its message processing loop (insensitive actor pattern applied where needed)
- All coordination uses message passing (no shared mutable state between actors)
- Message delivery is guaranteed but ordering is not assumed
- Addresses are treated as first-class data for capability routing
- Each actor specifies replacement behavior for every message type
- Actor creation happens dynamically based on computation needs
- Failure isolation prevents cascading failures across actor boundaries
- System topology can reconfigure at runtime without central registry
- Composition behavior verified, not just isolated component behavior
- Synchronous operations modeled as request-reply message pairs
NOT-FOR Boundaries
This skill should NOT be used for:
- Simple sequential computations → use functional programming instead
- Systems where shared memory is physically required → use lock-based concurrency patterns
- Real-time systems with hard timing constraints → use synchronous message passing with formal timing analysis
- Mathematical computations without coordination → use pure functional approaches
Delegate to other skills when:
- Implementing specific actor frameworks → use platform-specific implementation guides
- Performance tuning actor systems → use [performance-optimization] skill
- Formal verification of actor properties → use [formal-methods] skill
- Database design for actor persistence → use [data-architecture] skill
Common misconceptions about scope:
- Actors are not just "objects with async methods" - they require the full 3-tuple response
- Actor model is not just for distributed systems - applies to any concurrent computation
- Not primarily a performance optimization - it's a correctness and compositionality framework