Making your agent optimizable¶

Optimizers can only tune the parts of your agent that you expose to them. This page shows how to declare that surface in your Koog strategy so the optimizers have something to work on.

What gets optimized¶

An optimizer sees your agent as a set of tunable modules — the same idea as a DSPy module: a named unit with an instruction (and optional demonstrations) that the optimizer is free to rewrite. There are two levels:

The strategy system prompt — the __strategy__ module. This is the system prompt you pass when you build the agent. Every agent has it, so even a plain singleRunStrategy() agent (like the one in getting-started.md) is already optimizable: an optimizer can rewrite its system prompt, attach few-shot demonstrations, or append a learned playbook.
Optimizable subgraphs — finer-grained, per-step modules inside your strategy (the closest analogue to a DSPy module). When a single system prompt is too coarse — your agent reasons in stages, and each stage needs its own focused instruction — you declare those stages as optimizable subgraphs. The optimizer then tunes each one independently.

GEPA, MIPROv2, and BootstrapFewShot optimize the strategy system prompt plus every optimizable subgraph you declare. ACE currently optimizes only the strategy-level context (it curates a playbook for the __strategy__ module); per-subgraph optimization for ACE is a work in progress.

You don't need subgraphs to start. Reach for them when one prompt can't capture everything the agent must get right.

`optimizableSubgraphWithTask`¶

Declare an optimizable subgraph inside a strategy { } block with property delegation (by). It is the optimization-aware counterpart of Koog's subgraphWithTask: a subgraph that runs a single task whose instruction and few-shot demonstrations an optimizer can replace without you touching the graph.

import ai.koog.agents.optimization.core.optimizableSubgraphWithTask

val analyze by optimizableSubgraphWithTask<String, String>(
    optimizableInstruction = "Look at the customer message.",
    freshHistory = true,
) { instruction, message -> "Customer message:\n$message" }

The key parameters:

optimizableInstruction — the starting instruction. This is what runs before any optimization; an optimizer may replace it with a better one. Start vague or start with your best guess — either way the optimizer evolves it from here.
name — the module's name, defaulting to the delegated property name (analyze above). It must be unique within the strategy: the optimization artifact uses it as a lookup key.
freshHistory — when true, the subgraph starts with an empty conversation history. The resolved instruction is placed as a fresh system message and the task query follows it. This is a new parameter we added to Koog's subgraph(...) in the fork (expected to land upstream); it lets you (a) isolate a subgraph's prompt from the rest of the conversation, (b) give each subgraph its own clean, independently optimizable system message, and (c) keep prompts focused and cheaper for classification/routing-style steps that shouldn't inherit prior chatter.
defineTask { instruction, input -> ... } — builds the user message for this subgraph from the (resolved) instruction and the subgraph input, returning a String. The instruction argument already reflects any optimization; you usually compose your query from input and let the system message carry the instruction.

Few-shot knobs fewShotPromptType and demonstrationFormat control how learned demonstrations are inserted into the prompt (as message history vs. a single string, compact vs. detailed); leave them null to inherit sensible defaults. As a brief aside, the tool/LLM knobs — toolSelectionStrategy, llmModel, llmParams, runMode, assistantResponseRepeatMax, responseProcessor — let a subgraph use its own tools or model; they behave exactly as in a regular subgraphWithTask.

A worked example: a support-ticket router¶

This mirrors the runnable OptimizeSupportRouterExample. The agent is a two-stage router with two optimizable subgraphs:

nodeStart(message) -> remember -> [analyze] -> pack -> [route] -> nodeFinish(code)

Both subgraphs start from deliberately vague instructions that say nothing about the allowed routing codes or the required output format.

import ai.koog.agents.core.agent.entity.AIAgentGraphStrategy
import ai.koog.agents.core.agent.entity.createStorageKey
import ai.koog.agents.core.dsl.builder.node
import ai.koog.agents.core.dsl.builder.strategy
import ai.koog.agents.optimization.core.optimizableSubgraphWithTask

private val origMessageKey = createStorageKey<String>("orig-support-message")

private fun supportRouterStrategy(): AIAgentGraphStrategy<String, String> =
    strategy("support-router") {
        val remember by node<String, String>("remember") { message ->
            storage.set(origMessageKey, message)
            message
        }

        // Optimizable subgraph #1: ambiguous; doesn't say what to extract or why.
        val analyze by optimizableSubgraphWithTask<String, String>(
            optimizableInstruction = "Look at the customer message.",
            freshHistory = true,
        ) { _, message -> "Customer message:\n$message" }

        val pack by node<String, String>("pack") { analysis ->
            val message = storage.getValue(origMessageKey)
            "Customer message:\n$message\n\nNotes:\n$analysis"
        }

        // Optimizable subgraph #2: ambiguous; no codes, no output-format requirement.
        val route by optimizableSubgraphWithTask<String, String>(
            optimizableInstruction = "Decide how this request should be handled.",
            freshHistory = true,
        ) { _, packed -> packed }

        edge(nodeStart forwardTo remember)
        edge(remember forwardTo analyze)
        edge(analyze forwardTo pack)
        edge(pack forwardTo route)
        edge(route forwardTo nodeFinish)
    }

The dataset labels each ticket with the routing code it should produce, and the metric is an exact match on that code:

val dataset: TrainSet<String, String> = listOf(
    TrainSetItem("I was charged twice for my subscription this month.", "BILLING"),
    TrainSetItem("The app crashes every time I open the settings page.", "TECH"),
    TrainSetItem("Do you have a referral program for existing customers?", "OTHER"),
    // ...
)
val metric: (TrainSetItem<String, String>, String) -> Double = { item, answer ->
    if (answer.trim() == item.itemLabel) 1.0 else 0.0
}

The payoff. With the vague baseline instructions the agent has no idea the answer must be a bare code, so it replies with prose like "This should go to billing." and scores ~0 on the exact-match metric. An optimizer — GEPA here — reflects on the failure traces (which carry the expected labels), and evolves both subgraph instructions until analyze extracts the right signals and route emits exactly BILLING / TECH / OTHER. The learned instructions are saved to the optimization artifact and reapplied when you load the optimized agent.

You wire the optimizer to this agent exactly as in getting-started.md: build a trainingSession, call optimizer.train(session), then optimizer.loadOptimizedAgent(agent). See optimizers.md for how to pick and configure GEPA (or any other optimizer); we don't repeat that here.

Tips and gotchas¶

Names are artifact keys. Each optimizable subgraph must have a unique name within its strategy. The artifact maps instructions and demonstrations to subgraphs by name, so duplicates silently cause the wrong instruction to land on the wrong subgraph. Let names default to the property name and keep those distinct, or pass an explicit name.
freshHistory = true suits classification-style subtasks. When a subgraph should decide something from its input alone — classify, route, extract — a fresh history keeps it focused and makes its prompt cleanly optimizable. Leave it false when the step genuinely needs the prior conversation.
The two levels compose. Strategy-level and subgraph-level optimization are not exclusive — an optimizer tunes the __strategy__ system prompt and every subgraph in the same run. Start with just the strategy prompt; add optimizable subgraphs as you find steps that need their own instructions.

Next¶

optimizers.md — choose and configure an optimizer (GEPA, MIPROv2, BootstrapFewShot, ACE).
getting-started.md — the minimal end-to-end loop: build agent, dataset, metric, trainingSession, train, and run the optimized agent.