Use Functional Agent Skill

Process steps in order. Do not skip ahead.

Step 1 — Select the Agent Shape

• Use FunctionalAIAgent for a suspending Kotlin body with direct LLM requests.
• Use GraphAIAgent for explicit nodes, edges and subgraphs.
• Use PlannerAIAgent for runtime planning.

Proceed immediately to Step 2.

Step 2 — Create the Functional Strategy

Pass functionalStrategy to the top-level AIAgent factory. Supply the caller's executor; retain responsibility for closing that executor after its agents finish.

import ai.koog.agents.core.agent.AIAgent
import ai.koog.agents.core.agent.FunctionalAIAgent
import ai.koog.agents.core.agent.functionalStrategy
import ai.koog.prompt.executor.clients.openai.OpenAIModels
import ai.koog.prompt.executor.model.PromptExecutor
import ai.koog.prompt.message.MessagePart

fun formalEnglishAgent(executor: PromptExecutor): FunctionalAIAgent<String, String> =
    AIAgent<String, String>(
        promptExecutor = executor,
        llmModel = OpenAIModels.Chat.GPT4o,
        systemPrompt = "Rewrite the input in formal English. Return only the rewrite.",
        strategy = functionalStrategy {
            val response = requestLLM(it)
            response.parts.filterIsInstance<MessagePart.Text>()
                .joinToString("\n") { part -> part.text }.trim()
        },
    )

suspend fun rewrite(executor: PromptExecutor, input: String): String {
    val agent = formalEnglishAgent(executor)
    return try {
        agent.run(input)
    } finally {
        agent.close()
    }
}

The strategy receiver is AIAgentFunctionalContext. requestLLM adds the input to the prompt and requests a response. Read text from MessagePart.Text parts.

Proceed immediately to Step 3.

Step 3 — Check Feature Compatibility

• Install features through the factory's FunctionalAIAgent.FeatureContext.
• Confirm each selected feature accepts that context; compile its installation.
• Expect no graph node or edge events from this agent shape.
• For tools, register a ToolRegistry and implement a bounded loop using getToolCalls, executeTools and sendToolResults.
• Use a graph strategy when node-specific checkpoints or graph transitions are required.

Proceed immediately to Step 4.

Step 4 — Verify the Agent

Run ./gradlew build. Exercise rewrite with a mock executor returning a known text response; verify the trimmed rewrite and agent cleanup. Exercise any tool loop's termination bound and registered tools separately. If compilation or a behavior check fails, correct the factory, imports or strategy and repeat the checks.

Finish here.