Skip to content
Strands Harness
PYPython SDKTSTypeScript SDK
Strands
EVEvalsSHShell
Organizations
strands-agentsstrands-labs

Harness Optimizer

Harness Optimizer is a framework for optimizing an LLM agent’s harness through Formulas. The core idea: enhance the agent dynamically with tunable Formulas (for example, system prompts), then improve those Formulas with optimizers based on collected agent rollout trajectories.

from strands import Agent
from strands_harness_optimizer.formulas import SystemPromptFormula
from strands_harness_optimizer.adapters import apply_formulas_on_strands_agent


# Create a Formula
formula = SystemPromptFormula(system_prompt="You are a helpful assistant.")


# Attach to a Strands agent
agent = Agent(model=model)
apply_formulas_on_strands_agent(agent, [formula])


# Get / update parameters (e.g. after optimization)
params = formula.get_tunable_params()
formula.update_params({"system_prompt": "You are an expert coding assistant."})

Getting started

Section titled “Getting started”
Terminal window
pip install strands-harness-optimizer

How it works

Section titled “How it works”

The training loop is a small set of composable interfaces:

  1. DataLoader yields batches of task samples from a Dataset.
  2. AgentRolloutEngine executes the agent on each sample using the current Formula parameters, producing rollouts. An Adapter (apply_formulas_on_strands_agent) bridges Formula parameters to the agent.
  3. RewardFunction scores each rollout.
  4. Rollouts, data, and rewards are collected into a batch.
  5. FormulaOptimizer analyzes the batch to propose new Formula parameters, following PyTorch’s pattern: add_rollouts(), add_rewards(), step(), zero().
  6. The Formula updates its parameters and the loop repeats.

Built-in optimizers include a ContrastiveReflectionOptimizer, and rollout engines cover local execution and AgentCore.

Design decisions

Section titled “Design decisions”
  • Dict-based data throughout - no wrapper classes for context or results.
  • Minimal dependencies - core depends on strands-agents, strands-agents-tools, jinja2, and botocore.
  • PyTorch-style Dataset / DataLoader reused from stdlib, with no PyTorch dependency.
  • Minimal Trainer - the built-in training loop is just two nested for-loops; bring your own if you prefer.
Section titled “Links”