takeokunn/Parallelization Patterns

<purpose> Provide standardized patterns for parallel agent execution, timeout configuration, retry policies, and consensus mechanisms shared across orchestrators and agents. </purpose> <tools> <tool name="parallelization_template"> <description>Standard parallel execution capability and strategy patterns</description> <use_case>Include in agents and commands for consistent parallel execution</use_case> </tool> </tools> <concepts> <concept name="parallel_safety"> <description>Classification of agents by their parallel execution safety</description> <example> read_only: Safe to run with any other agent (explore, validator) analysis: Safe with other analysis agents (design, database, performance) execution: May modify local state, requires coordination (security, test, devops) orchestration: Manages sub-agents (commands, CLAUDE.md) </example> </concept> <concept name="timeout_tiers"> <description>Standard timeout values by agent type</description> <example> Fast read-only (explore): 180000ms (3 min) Medium analysis (design, database, quality-assurance): 240000ms (4 min) Deep analysis (security, performance, test): 300000ms (5 min) Complex operations (code-quality, devops, git, docs): 300000ms (5 min) Orchestration (commands, CLAUDE.md): 300000ms (5 min) </example> </concept> </concepts> <patterns> <pattern name="parallelization_readonly"> <description>Parallelization pattern for read-only agents (explore, validator)</description> <example> <parallelization> <capability> <parallel_safe>true</parallel_safe> <read_only>true</read_only> <modifies_state>none</modifies_state> </capability> <execution_strategy> <max_parallel_agents>16</max_parallel_agents> <timeout_per_agent>180000</timeout_per_agent> </execution_strategy> <safe_with> <agent>all read-only agents</agent> </safe_with> <conflicts_with /> </parallelization> </example> </pattern> <pattern name="parallelization_analysis"> <description>Parallelization pattern for analysis agents (design, database, code-quality, performance)</description> <example> <parallelization> <capability> <parallel_safe>true</parallel_safe> <read_only>true</read_only> <modifies_state>none</modifies_state> </capability> <execution_strategy> <max_parallel_agents>16</max_parallel_agents> <timeout_per_agent>240000</timeout_per_agent> </execution_strategy> <safe_with> <agent>other analysis agents</agent> </safe_with> <conflicts_with /> </parallelization> </example> </pattern> <pattern name="parallelization_execution"> <description>Parallelization pattern for execution agents (security, test, devops)</description> <example> <parallelization> <capability> <parallel_safe>true</parallel_safe> <read_only>false</read_only> <modifies_state>local</modifies_state> </capability> <execution_strategy> <max_parallel_agents>16</max_parallel_agents> <timeout_per_agent>300000</timeout_per_agent> </execution_strategy> </parallelization> </example> </pattern> <pattern name="parallelization_orchestration"> <description>Parallelization pattern for orchestrators (CLAUDE.md, commands)</description> <example> <parallelization> <capability> <parallel_safe>true</parallel_safe> <read_only>false</read_only> <modifies_state>orchestration</modifies_state> </capability> <execution_strategy> <max_parallel_agents>16</max_parallel_agents> <timeout_per_agent>300000</timeout_per_agent> <parallel_groups> <group id="investigation" agents="explore,design,database,performance" independent="true"/> <group id="quality" agents="code-quality,security,test" independent="true"/> <group id="review" agents="quality-assurance,docs" independent="true"/> <group id="validation" agents="validator" independent="false" depends_on="investigation,quality,review"/> </parallel_groups> </execution_strategy> </parallelization> </example> </pattern> <pattern name="retry_policy"> <description>Standard retry policy for agent execution failures</description> <example> <retry_policy> <max_retries>2</max_retries> <retry_conditions> <condition>Agent timeout</condition> <condition>Partial results returned</condition> <condition>Confidence score below 60</condition> </retry_conditions> <fallback_strategy> <action>Use alternative agent from same parallel group</action> </fallback_strategy> </retry_policy> </example> </pattern> <pattern name="consensus_thresholds"> <description>Standard consensus thresholds for multi-agent validation</description> <example> <consensus_thresholds> <threshold level="high" value="0.9" action="Auto-accept without review"/> <threshold level="medium" value="0.7" action="Accept with note"/> <threshold level="low" value="0.5" action="Flag for user review"/> <threshold level="conflict" value="0.5" action="Block, require user decision"/> </consensus_thresholds> </example> </pattern> <pattern name="agent_weights"> <description>Standard agent weights for consensus mechanism</description> <example> <weights> <agent name="explore" weight="1.0"/> <agent name="design" weight="1.2"/> <agent name="database" weight="1.2"/> <agent name="performance" weight="1.2"/> <agent name="code-quality" weight="1.1"/> <agent name="security" weight="1.5"/> <agent name="test" weight="1.1"/> <agent name="docs" weight="1.0"/> <agent name="quality-assurance" weight="1.3"/> <agent name="devops" weight="1.1"/> <agent name="validator" weight="2.0"/> </weights> </example> </pattern>

<decision_tree name="timeout_selection"> <question>What type of agent is this?</question> <branch condition="Fast read-only (explore)">180000ms</branch> <branch condition="Medium analysis (design, database, quality-assurance)">240000ms</branch> <branch condition="Deep analysis (security, performance, test)">300000ms</branch> <branch condition="Complex operations (code-quality, devops, git, docs)">300000ms</branch> <branch condition="Orchestration (commands, CLAUDE.md)">300000ms</branch> </decision_tree>

<decision_tree name="parallelization_selection"> <question>What does this agent modify?</question> <branch condition="Nothing (read-only)">Use parallelization_readonly</branch> <branch condition="Analysis only">Use parallelization_analysis</branch> <branch condition="Local state">Use parallelization_execution</branch> <branch condition="Orchestrates sub-agents">Use parallelization_orchestration</branch> </decision_tree> </patterns>

<best_practices> <practice priority="critical">Select appropriate parallelization pattern based on agent capabilities</practice> <practice priority="critical">Select timeout based on agent type following decision_tree</practice> <practice priority="high">Group independent agents in parallel_groups for concurrent execution</practice> <practice priority="high">Define depends_on for validation agents that require prior results</practice> <practice priority="medium">Use retry_policy for critical operations</practice> </best_practices>

<rules priority="critical"> <rule>Read-only agents must use parallelization_readonly or parallelization_analysis</rule> <rule>Agents that modify state must use parallelization_execution</rule> <rule>Orchestrators must use parallelization_orchestration</rule> </rules> <rules priority="standard"> <rule>Use timeout_selection decision tree for consistent timeout values</rule> <rule>Apply retry_policy for operations that may fail transiently</rule> <rule>Use consensus_thresholds for multi-agent validation scenarios</rule> </rules> <constraints> <must>Match parallelization pattern to agent's state modification behavior</must> <must>Use standard timeout tiers from timeout_selection</must> <must>Define parallel_groups with correct dependency chains</must> <avoid>Running write agents in parallel without coordination</avoid> <avoid>Using non-standard timeout values</avoid> </constraints>

<related_skills> <skill name="core-patterns">Base templates for error escalation, decision criteria, enforcement</skill> <skill name="workflow-patterns">Output formats, reflection checkpoints, self-evaluation</skill> </related_skills>

<related_agents> <agent name="validator">Validate that parallel task results are consistent and non-conflicting</agent> </related_agents>