examples/agentic-system-prompting/README.md

Agent Prompt Optimizer

The Problem with Manual System Prompting

• Time-consuming iteration cycles of trial and error
• Prompt drift and regression as you patch edge cases
• No systematic learning from agent failures
• Knowledge stays in your head instead of in the prompt
• Hard to manage as prompts scale

The Solution

ACE (Agentic Context Engine) automatically optimizes your agent's system prompt by learning from execution. It observes agent runs, analyzes what strategies worked and what failed, then generates actionable insights for the system prompt.

Traces / Conversations → ACE → Prompt Suggestions

You put in past traces or conversations. ACE handles agentic system prompting by learning from mistakes. You receive improved system prompt suggestions.

How it works:

0. Prepare your data - Export/convert your agent conversations to .md or .toon files and place them in a directory. To convert JSON to TOON, use the included convert.py script or the toon library directly. The more detailed your traces, the better the insights.

1. ReplayAgent - Simulates an agent for offline learning from your trace/conversation

2. Reflector - Analyzes each conversation to identify what worked, what failed, and why

3. SkillManager - Transforms reflections into atomic, actionable prompt strategies/insights

4. Deduplicator - Consolidates similar strategies/insights using embeddings to keep the output clean

5. Skillbook - Output file stores all prompt strategies/insights in a human-readable format you can review and implement

The output is a human-readable skillbook where each insight contains:

• Prompt suggestion - The recommended text to add to your system prompt
• Justification - Why this change would help based on the analysis
• Evidence - What actually happened in the trace that led to this insight You review each suggestion and decide what to copy into your system prompt. ACE may even suggest strategies that contradict your current prompt when it identifies flaws in the original design.

Setup

Installation

pip install ace-framework

Or for development:

git clone https://github.com/kayba-ai/agentic-context-engine
cd agentic-context-engine
uv sync
uv pip install -e .  # Required to run examples

API Keys

Requirements:

• LLM API key for analysis (e.g., OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.)
• OPENAI_API_KEY for deduplication (uses OpenAI embeddings)

Create a .env file in the project root:

# Required for analysis (choose one)
OPENAI_API_KEY=your-openai-key
# OR
ANTHROPIC_API_KEY=your-anthropic-key

# Required for deduplication (uses OpenAI embeddings)
OPENAI_API_KEY=your-openai-key

Implementation

Agentic System Prompting = ACE Offline Adapter

Process past traces and conversations in batch to generate insights without the agent running.

Use case: Periodic automated system prompt revision. Feed historical data, let ACE analyze patterns, then have a human review and choose what to implement.

Quick Start (CLI)

# Basic usage
python agentic_system_prompting.py /path/to/traces

# With options
python agentic_system_prompting.py /path/to/traces --model gpt-4o --epochs 2
python agentic_system_prompting.py /path/to/traces --input-skillbook existing.json
python agentic_system_prompting.py /path/to/traces --output-dir ./results --threshold 0.8

CLI Options:

• traces_dir - Required: path to directory containing .md or .toon trace files
• -m, --model - LLM model for analysis (default: claude-haiku-4-5-20251001)
• -e, --epochs - Number of training epochs (default: 1)
• -t, --threshold - Deduplication similarity threshold 0.0-1.0 (default: 0.7)
• -i, --input-skillbook - Continue learning from an existing skillbook
• -o, --output-dir - Output directory for results (default: script directory)

Outputs:

• skillbook_{timestamp}.json - The learned skillbook
• skills_{timestamp}.md - Human-readable skills grouped by section
• external_agent_injection_{timestamp}.txt - Ready-to-inject prompt text for external agents

Python API

from ace import (
    Skillbook,
    Sample,
    OfflineACE,
    Reflector,
    SkillManager,
    ReplayAgent,
    SimpleEnvironment,
)
from ace.llm_providers.litellm_client import LiteLLMClient, LiteLLMConfig
from ace.prompts_v3 import PromptManager

# 1. Initialize LLM client
config = LiteLLMConfig(
    model="claude-sonnet-4-5-20250929",
    max_tokens=8192,
    temperature=0.1,
)
llm = LiteLLMClient(config=config)
prompt_mgr = PromptManager()

# 2. Create ACE components
skillbook = Skillbook()
agent = ReplayAgent()  # Dummy agent that replays conversations
reflector = Reflector(llm=llm, prompt_template=prompt_mgr.get_reflector_prompt())
skill_manager = SkillManager(llm=llm, prompt_template=prompt_mgr.get_skill_manager_prompt())

# 3. Load past conversations as samples
samples = [
    Sample(
        question="Your task description here",
        context="The full conversation/trace content",
        ground_truth="",  # Empty for analysis tasks
        metadata={"source": "conversation_1"}
    ),
    # ... more historical data
]

# 4. Create adapter and run
environment = SimpleEnvironment()
adapter = OfflineACE(
    skillbook=skillbook,
    agent=agent,
    reflector=reflector,
    skill_manager=skill_manager,
)
results = adapter.run(samples, environment, epochs=1)

# 5. Review and save generated skills
print(adapter.skillbook.as_prompt())
adapter.skillbook.save_to_file("offline_adapter_skillbook.json")

Tip: Enable deduplication to automatically consolidate similar skills during learning. This keeps the skillbook clean.

from ace import DeduplicationConfig

dedup_config = DeduplicationConfig(
    enabled=True,
    similarity_threshold=0.85,
)

adapter = OfflineACE(
    skillbook=skillbook,
    agent=agent,
    reflector=reflector,
    skill_manager=skill_manager,
    dedup_config=dedup_config,
)

Async Mode

For large batches, enable async learning so the Reflector and SkillManager process in the background:

adapter = OfflineACE(
    skillbook=skillbook,
    agent=agent,
    reflector=reflector,
    skill_manager=skill_manager,
    async_learning=True,
    max_reflector_workers=3,
)

results = adapter.run(samples, environment)

Checkpoints

Save skillbook periodically during long training runs:

results = adapter.run(
    samples=samples,
    environment=environment,
    epochs=3,
    checkpoint_interval=10,  # Save every 10 samples
    checkpoint_dir="./checkpoints",
)

Agentic Prompting at Runtime = Online Adapter

Fully autonomous self-improving agents at runtime. The agent learns from every interaction, generates insights, and injects them into future contexts automatically - no manual intervention required.

Use case: Continuous improvement in production where agents get better with every run.

See the Quick Start Guide for setup instructions.

• LiteLLM: examples/litellm/ - Make a new agent that self-learns
• LangChain: examples/langchain/ - Wrap your existing agent with self-improving

FAQ

Can I combine manual prompts with ACE skills? Yes. ACE skills complement your base prompts. Start with manual prompts and let ACE build domain-specific expertise on top.

What if ACE suggests something that contradicts my system prompt? Review it. ACE may have identified a flaw in your original design. The skillbook is human-readable JSON - you decide what to keep.

Can I share skills between agents? Yes. Skillbooks are portable JSON files with human readable text.

Next Steps

• Explore examples in this repository
• Read the main documentation
• Join our Discord for tips and support