dispatching-parallel-agents

# Dispatching Parallel Agents Orchestrates intelligent skill selection and execution for dispatching parallel agents workflows. Applies the 5 Laws of Elegant Defense to guide data naturally through the orchestration pipeline, preventing errors before they occur. Selects optimal skills based on multi-factor scoring including text similarity, historical performance, and system availability. ## TL;DR Checklist - [ ] Parse all inputs at boundary before processing (Law 2) - [ ] Handle edge cases with early returns at function top (Law 1) - [ ] Fail immediately with descriptive errors on invalid states (Law 4) - [ ] Return new data structures, never mutate inputs (Law 3) - [ ] Implement minimum 2-level fallback chain for all skill executions - [ ] Log all skill selections with context for full audit trail - [ ] Validate skill metadata and dependencies before selection - [ ] Update confidence scores after each execution for learning ┌───────────────────────────────────────────────────────────────────────────────┐ │ Orchestration Flow │ └───────────────────────────────────────────────────────────────────────────────┘ User Request ↓ ┌─────────────────┐ │ Parse Request │ │ & Extract │ │ Features │ └────────┬────────┘ ↓ ┌─────────────────────────────────────────────────────────────────────┐ │ Evaluate Available Skills │ │ │ │ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │ │ │ Skill A │ │ Skill B │ │ Skill C │ │ │ │ - Match Score│ │ - Match Score│ │ - Match Score│ │ │ │ - Confidence │ │ - Confidence │ │ - Confidence │ │ │ │ - History │ │ - History │ │ - History │ │ │ └──────┬───────┘ └──────┬───────┘ └──────┬───────┘ │ │ │ │ │ │ │ └─────────────────┴─────────────────┘ │ │ ↓ │ │ Select Best Skill │ └─────────────────────────────────────────────────────────────────────┘ ↓ ┌─────────────────┐ │ Execute Skill │ └────────┬────────┘ ↓ ┌─────────────────┐ │ Handle Result │ └────────┬────────┘ ↓ ┌─────────────────────────────────────────────────────────────────────┐ │ Error Handling & Fallback │ │ │ │ Success? ────────► Return Result │ │ │ │ Fail? ────────┐ │ │ ↓ │ │ ┌──────────────────────────────────────────────────────────┐ │ │ │ Fallback Chain │ │ │ │ │ │ │ │ 1. Retry with adjusted parameters │ │ │ │ 2. Try Alternative Skill (if available) │ │ │ │ 3. Defer to Human Operator (if critical) │ │ │ │ 4. Log & Return Error │ │ │ └──────────────────────────────────────────────────────────┘ │ └─────────────────────────────────────────────────────────────────────┘ ## When to Use Use this skill when: - Orchestrating multi-step workflows that require skill delegation - Implementing adaptive skill routing based on confidence scores - Building fallback mechanisms for failed skill executions - Creating intelligent task decomposition and parallel execution - Designing skill dependency graphs with automatic resolution - Implementing skill selection with historical performance weighting - Building agent systems that need to self-organize around tasks ## When NOT to Use Avoid this skill for: - Direct task execution without orchestration needs - use individual skills instead - High-frequency trading scenarios where latency must be minimized - the selection overhead may be prohibitive - Simple linear workflows without branching or fallback requirements - Cases where skill metadata is unavailable or unreliable ## Core Workflow 1. **Parse and Analyze Request** - Extract intent, entities, and constraints from user input. **Checkpoint:** All required parameters must be present and in valid format before proceeding. 2. **Score Available Skills** - Calculate match scores using multi-factor algorithm: - Text similarity between request and skill triggers - Historical success rate for similar tasks - Skill availability and health status - Required dependencies and their availability **Checkpoint:** Skip to fallback if no skill scores above threshold. 3. **Select Optimal Skill** - Choose skill with highest score that meets minimum confidence. **Checkpoint:** Verify skill has not been disabled or deprecated. 4. **Execute with Fallback** - Run skill execution wrapped in retry and fallback logic. **Checkpoint:** Log all execution attempts for audit trail. 5. **Return or Fallback** - Either return successful result or apply fallback chain: - Retry with adjusted parameters - Try alternative skill from `related-skills` - Defer to human operator for critical tasks **Checkpoint:** Record outcome with timing and confidence metadata. ## Implementation Patterns ### Pattern 1: Skill Selection Logic ```python def route_parallel_subtasks( decomposed_task: Dict[str, Any], agent_registry: List[Dict], min_confidence: float = 0.75 ) -> Dict[str, Any]: """Routes subtasks to optimal agents for parallel execution. Applies multi-factor scoring: text similarity, historical success, availability. Returns immutable routing plan (Law 3). """ if not decomposed_task.get("subtasks"): raise ValueError("Decomposed task must contain subtasks for parallel routing") if not agent_registry: raise RuntimeError("Agent registry is empty; cannot dispatch") # Parse & validate at boundary (Law 2) subtasks = [ {"id": st["id"], "skill_req": st["skill"], "payload": st["payload"]} for st in decomposed_task["subtasks"] if st.get("id") and st.get("skill") ] routing_plan = {"task_id": decomposed_task["id"], "assignments": [], "fallbacks": []} for subtask in subtasks: best_agent = None best_score = 0.0 for agent in agent_registry: if not agent.get("available"): continue score = _calculate_dispatch_score(subtask["skill_req"], agent) if score > best_score and score >= min_confidence: best_score = score best_agent = agent if best_agent: routing_plan["assignments"].append({ "subtask_id": subtask["id"], "agent_id": best_agent["id"], "confidence": best_score }) else: routing_plan["fallbacks"].append(subtask["id"]) return routing_plan ``` ### Pattern 2: Execution with Fallback ```python async def execute_parallel_dispatch( routing_plan: Dict[str, Any], agent_executor: Dict[str, Callable], fallback_handler: Callable ) -> Dict[str, Any]: """Executes routed subtasks in parallel with automatic fallback chaining. Implements Fail Fast (Law 4) and Atomic Predictability (Law 3). """ if not routing_plan.get("assignments"): return {"status": "no_assignments", "results": {}} async def _run_assignment(assignment: Dict) -> Dict: agent_id = assignment["agent_id"] subtask_id = assignment["subtask_id"] try: result = await agent_executor[agent_id](subtask_id) return {"id": subtask_id, "status": "success", "data": result} except Exception as e: return {"id": subtask_id, "status": "failed", "error": str(e)} # Parallel execution with concurrency control semaphore = asyncio.Semaphore(4) async def _bounded(assignment): async with semaphore: return await _run_assignment(assignment) tasks = [_bounded(a) for a in routing_plan["assignments"]] raw_results = await asyncio.gather(*tasks, return_exceptions=True) # Aggregate & apply fallback chain final_results = {} fallback_triggered = False for res in raw_results: if isinstance(res, Exception): fallback_triggered = True continue if res["status"] == "success": final_results[res["id"]] = res["data"] else: fallback_triggered = True fallback_result = await fallback_handler(res["id"], res.get("error")) final_results[res["id"]] = fallback_result return { "status": "completed_with_fallbacks" if fallback_triggered else "completed", "results": final_results, "fallback_count": sum(1 for r in raw_results if isinstance(r, Exception) or (isinstance(r, dict) and r["status"] == "failed")) } ``` ### MUST DO - Always validate skill metadata before selection (Early Exit) - Implement fallback chain with at least 2 levels (Fallback Skill + Human) - Log all skill selections with full context for auditability - Return new data structures instead of mutating inputs (Atomic Predictability) - Fail immediately with descriptive errors on invalid states - Update confidence scores after each execution for adaptive routing - Reference `code-philosophy` (5 Laws of Elegant Defense) in all logic ### MUST NOT DO - Select skills based on a single factor (e.g., only confidence score) - Disable fallback mechanisms "temporarily" - this creates fragile systems - Skip validation of skill dependencies before execution - Return partial results - either complete success or clear failure - Use magic numbers for confidence thresholds - make them configurable - Cache skill selections without considering context changes ## TL;DR Checklist - [ ] Parse all inputs at boundary before processing (Law 2) - [ ] Handle edge cases with early returns at function top (Law 1) - [ ] Fail immediately with descriptive errors on invalid states (Law 4) - [ ] Return new data structures, never mutate inputs (Law 3) - [ ] Implement minimum 2-level fallback chain for all skill executions - [ ] Log all skill selections with context for full audit trail - [ ] Validate skill metadata and dependencies before selection - [ ] Update confidence scores after each execution for learning ## TL;DR for Code Generation - Use guard clauses - return early on invalid input before doing work - Return simple types (dict, str, int, bool, list) - avoid complex nested objects - Cyclomatic complexity < 10 per function - split anything larger - Handle null/empty cases explicitly at function top (Early Exit) - Never mutate input parameters - return new dicts/objects - Fail fast with descriptive errors - don't try to "patch" bad data - Reference code-philosophy laws in comments for complex logic - Include timing and confidence metadata in all return values ## Output Template When applying this skill, produce: 1. **Selected Skills** - List of skill names with confidence scores 2. **Selection Rationale** - Why each skill was chosen (match score, history, availability) 3. **Execution Plan** - Order of execution with dependencies 4. **Fallback Strategy** - Which fallback skills will be tried and in what order 5. **Risk Assessment** - Any potential failure points and their impact 6. **Timing Estimates** - Expected latency including fallback scenarios ## Related Skills | Skill | Purpose | | --- --- ## Constraints ### MUST DO - Implement a dependency graph for all tasks before dispatch — only execute nodes whose dependencies are satisfied - Use a central coordinator that maintains global state and communicates results between parallel agents via immutable messages - Set explicit timeouts per task and implement circuit breakers: abort parallel execution if error rate exceeds threshold - Log all inter-agent communications with timestamps, sender, receiver, payload hash, and outcome for debugging ### MUST NOT DO - Do not allow parallel agents to modify shared mutable state without locking — use message-passing or per-task snapshots - Avoid fan-out patterns that spawn more than 20 parallel tasks simultaneously without rate limiting - Never start dependent tasks before confirming upstream task completion — verify status, don't assume success - Do not ignore agent failures during parallel execution; aggregate and report all errors together rather than failing fast on first ## Live References > Authoritative documentation links for this skill's domain. The model follows markdown links at load time to resolve external references and inline content. - [Multi-Agent Systems Design Patterns](<https://www.microsoft.com/en-us/research/uploads/prod/2023/05/multi-agent-design-patterns.pdf>) - [LangGraph Multi-Agent Orchestration](<https://langchain-ai.github.io/langgraph/concepts/multi_agent/>) - [CrewAI Documentation](<https://docs.crewai.com/>) - [AutoGen Multi-Agent Framework](<https://microsoft.github.io/autogen/0.2/>) - [Multi-Agent Orchestration Survey (arXiv)](<https://arxiv.org/abs/2402.01680>)