subagent-task

name	subagent-task
description	Guidance and protocols for spawning sub-agents via the task tool to handle complex, multi-step, or parallelizable tasks. You MUST load this skill when the task tool for invoking sub-agents is available.
license	MIT

subagent-task

Provides policies and examples for using the task tool to spawn sub-agents for specialized, parallel, or complex tasks.

WHEN TO USE

• Delegating complex, multi-step tasks to specialized project agents.
• Executing modular sub-tasks in parallel (e.g., retrieving facts, analyzing logs, validating plans).
• Preventing context clutter in the primary agent by offloading heavy reading or analysis to a sub-agent.
• Utilizing specific personas for targeted expertise.

WHEN NOT TO USE

• For trivial, single-step operations (e.g., running a quick ls or a simple git commit).
• When the overhead of explaining the task to a sub-agent outweighs doing it directly.
• If the primary agent is already a specialized agent perfectly suited for the entire task.

Core Principles

• Specialization: Delegate tasks to specialized project agents rather than attempting to handle broad contexts monolithically.
• Parallelization: Spawn multiple agents concurrently for independent sub-tasks (e.g., retrieving facts, analyzing logs, validating plans).
• Context Encapsulation: Spawning sub-agents prevents the primary agent's context from becoming cluttered. Synthesize results from sub-agents into the final response.

SUBAGENT DELEGATION POLICY

The use of the task tool and spawning sub-agents is permitted for complex, multi-step tasks, but delegation is limited to the project agents explicitly configured by this runtime.

• Allowed Delegation Targets: Use only the named project agents exposed by this runtime configuration.
• Built-in Subagents Disabled: Built-in explore and general subagents are not approved for this runtime and MUST NOT be used, even if a host tool still lists them.
• Maintain Context: Ensure that the primary agent remains the coordinator and synthesizes the results from sub-agents into the final response.
• Strategic Delegation: Delegate only when the task involves broad codebase analysis or independent sub-tasks that can be executed in parallel.

Common Pitfalls

• Vague Delegation Prompts: Sending ambiguous instructions to the sub-agent ("fix the code") without explicit context or success criteria, resulting in poor outcomes.
• Ignoring Sub-Agent Output: Failing to properly read and synthesize the results returned by the task tool before taking further action.
• Spawning the Wrong Agent Type: Delegating a task to an inappropriate agent (e.g., using programmer for DevOps automation instead of devops), leading to failed runs.

Example: Agent Delegation Protocol

flowchart TD
    %% --------------------------------------------------------
    %% Agent Delegation Protocol
    %% --------------------------------------------------------
    subgraph Flow_Agent_Delegation ["Agent Delegation Protocol"]
        direction TB
        Arch((Cogni AI Architect))
        Arch -->|task tool| DevOps[cogni-ai-devops<br/>Task automation, CI/CD, IaC]
        Arch -->|task tool| FactOps[cogni-ai-fact-ops<br/>Maintain facts & info consistency]
        Arch -->|task tool| Coder[cogni-ai-coder<br/>Spec-driven code implementation]
        Arch -->|task tool| Programmer[cogni-ai-programmer<br/>Algorithmic problem-solving]
        Arch -->|task tool| Developer[cogni-ai-developer<br/>Full-cycle product delivery]
        Arch -->|task tool| BrainOps[cogni-ai-brain-ops<br/>Gather facts, constraints, plan]
        Arch -->|task tool| GHOps[cogni-ai-github-ops<br/>Modify comments, issues, discussions]
        Arch -->|task tool| CodeRev[cogni-ai-code-reviewer<br/>PR analysis, quality, security]
        Arch -->|task tool| SecAud[cogni-ai-security-auditor<br/>Security audit & PR review]
        Arch -->|task tool| PlanRev[cogni-ai-plan-reviewer<br/>Architectural & plan validation]
        Arch -->|task tool| Tester[cogni-ai-tester<br/>Execute tests, verify behavior]
        Arch -->|task tool| C7Ops[cogni-ai-context7-ops<br/>Retrieve & filter documentation]
    end

Example: Delegation Scenarios

flowchart TD
    Trigger{Task Requirement}

    Trigger -->|Automation/CI/CD/IaC| DevOps["devops<br/>CI/CD and automation tasks"]
    Trigger -->|Complex task| Brain["brain-ops<br/>Brainstorm and architect plan"]
    Trigger -->|Docs updated| Docs["docs-editor<br/>Review and update documentation"]
    Trigger -->|GitHub writes| GHOps["github-ops<br/>Invoke GitHub write operations"]
    Trigger -->|Plan validation| PlanRev["plan-reviewer<br/>Validate plan plausibility"]
    Trigger -->|Python needed| Programmer["programmer<br/>Write and test Python code"]
    Trigger -->|Large diff| CodeRev["code-reviewer<br/>Final code quality review"]
    Trigger -->|Security audit/review| SecAud["security-auditor<br/>Perform security audit or review"]

    classDef scenario fill:#f9f,stroke:#333,stroke-width:2px
    class Docs,Brain,Programmer,GHOps,PlanRev,CodeRev,SecRev,DevOps scenario

Usage Patterns

• Always pass a clear description and a detailed prompt to the sub-agent.
• Provide the subagent_type argument to match the desired role. Ensure you check your available tools dynamically to find the currently supported list of project-specific agents rather than relying on a hardcoded list, as this functionality exists and is strongly recommended to be utilized for more specialized agents for relevant context.
• Ensure the primary agent acts as a coordinator, processing the task_result from each sub-agent before continuing the plan.
• If a sub-agent misbehaves (e.g., returning an unexpected reply) or fails to meet expectations, report this issue to the user with a clear explanation.

What to Avoid

• Do NOT use subagent_type for skills: The subagent_type parameter is reserved for agent types. Skills are loaded inside the sub-agent via the skill tool.

Related Skills

• critical-thinking: You MUST load this skill when deconstructing complex tasks for sub-agent delegation.
• gh: You MUST load this skill when working with the gh command and its subcommands.