Sub-Agent Delegation Rules

Platform-agnostic. The context-loss problem is the same on Claude Code and Copilot.

The universal rule

• Sub-agents / extensions / participants start with FRESH context. They do NOT inherit parent's skills, history, or accumulated state.
• Pass skills explicitly — every platform, every time.
• Validate the handoff. First sub-agent action: echo received skills. Fail loudly on mismatch.

Claude Code

• Inherit: CLAUDE.md, tool definitions, MCP servers.
• Pass skills: AgentDefinition(skills=[...]) or inline in Task tool prompt.
• context: fork is INVERTED (GH #20492) — creates blank, not copy.
• AskUserQuestion unavailable in sub-agents (GH #34592).

VS Code Copilot

• Agent mode persists within session BUT terminal commands and chat participants get fresh context.
• Scripts the agent writes that spawn sub-processes inherit NOTHING from Copilot.
• Pass skills: #file:path/to/SKILL.md in chat, or .github/copilot-instructions.md for always-on.

Koog (Kotlin/JVM) — no native skill registry

Koog 1.0.0-preview3's AIAgent(...) factory only takes a systemPrompt. There is no skills=[...] parameter. Native skill support is tracked at JetBrains/koog#1383.

Fallback while the native API is missing: read the installed skill files and inject their content into the sub-agent's systemPrompt at agent-creation time. Do NOT copy-paste skill content into your source — drift between the plugin repo and your code turns the demo's "skill handoff" beat into theatre.

private fun readSkillRules(pluginName: String): String {
    val tileDir = listOf(
        File(".tessl/tiles/jbaruch/$pluginName"),
        File("../../.tessl/tiles/jbaruch/$pluginName")
    ).firstOrNull { it.isDirectory }
        ?: error("Run `tessl install jbaruch/$pluginName` before invoking this sub-agent.")
    return File(tileDir, "rules").listFiles { f -> f.extension == "md" }
        ?.sortedBy { it.name }
        ?.joinToString("\n\n---\n\n") { it.readText() } ?: ""
}

val visionAgent = AIAgent(
    promptExecutor = PromptExecutor.builder().anthropic(apiKey).build(),
    llmModel = AnthropicModels.Sonnet_4,
    strategy = functionalStrategy<String, String> { input ->
        requestLLM(input).parts.filterIsInstance<MessagePart.Text>().joinToString("\n") { it.text }
    },
    systemPrompt = """
        You are the vision sub-agent.

        SKILL (jbaruch/face-recognition-calibration-djl):
        ${readSkillRules("face-recognition-calibration-djl")}

        ... task description ...
    """.trimIndent()
)

When Koog ships native skills (koog#1383), swap readSkillRules(...) for the native registration call. The same fallback pattern applies to any framework where sub-agents only accept a system prompt.

Cross-platform ground truth