AI agents don't learn from experience. They repeat the same mistakes every session, forget what worked, and ignore what failed. ACE is the open-source engine that adds a persistent learning loop. It also powers Kayba, the managed service that does this for your production agents automatically.

The agent claims a seahorse emoji exists. ACE reflects on the error, and on the next attempt, the agent responds correctly — without human intervention.

uv add ace-framework

Option A — Interactive setup (recommended):

ace setup            # Walks you through model selection, API keys, and connection validation

Option B — Manual configuration:

export OPENAI_API_KEY="your-key"    # or ANTHROPIC_API_KEY, or any of 100+ supported providers

Then use it:

from ace import ACELiteLLM

agent = ACELiteLLM(model="gpt-4o-mini")

# First attempt — the agent may hallucinate
answer = agent.ask("Is there a seahorse emoji?")

# Feed a correction — ACE extracts a strategy and updates the Skillbook
agent.learn_from_feedback("There is no seahorse emoji in Unicode.")

# Subsequent calls benefit from the learned strategy
answer = agent.ask("Is there a seahorse emoji?")

# Inspect what the agent has learned
print(agent.get_strategies())

No fine-tuning, no training data, no vector database.

-> Quick Start Guide | -> Setup Guide | -> Hosted API: Where Do Traces Come From?

ACE maintains a Skillbook — a persistent collection of strategies that evolves with every task. Three specialized roles manage the learning loop:

The Recursive Reflector is the key innovation: instead of summarizing traces in a single pass, it writes and executes Python code in a sandboxed environment to programmatically search for patterns, isolate errors, and iterate until it finds actionable insights.

flowchart LR
    Skillbook[(Skillbook)]
    Start([Task]) --> Agent[Agent]
    Agent <--> Environment[Environment]
    Environment -- Trace --> Reflector[Reflector]
    Reflector --> SkillManager[SkillManager]
    SkillManager -- Updates --> Skillbook
    Skillbook -. Strategies .-> Agent

All roles are backed by PydanticAI agents with structured output validation. PydanticAI routes to 100+ LLM providers through its LiteLLM integration, with native support for OpenAI, Anthropic, Google, Bedrock, Groq, and more.

Based on the ACE paper (Stanford & SambaNova) and Dynamic Cheatsheet.

uv add 'ace-framework[browser-use]'    # Browser automation
uv add 'ace-framework[langchain]'      # LangChain
uv add 'ace-framework[logfire]'        # Observability (auto-instruments PydanticAI)
uv add 'ace-framework[mcp]'            # MCP server for IDE integration
uv add 'ace-framework[deduplication]'  # Embedding-based skill deduplication

Have existing agent logs? Extract strategies from them directly:

from ace import ACELiteLLM

agent = ACELiteLLM(model="gpt-4o-mini")
agent.learn_from_traces(your_existing_traces)
print(agent.get_strategies())

-> Examples

tau2-bench by Sierra Research: airline domain tasks requiring tool use and policy adherence. Claude Haiku 4.5 agent, strategies learned on the train split with no reward signals, evaluated on the held-out test split.

pass^k = probability all k independent attempts succeed. ACE doubles consistency at pass^4 with 15 learned strategies.

ACE + Claude Code translated this library from Python to TypeScript with zero supervision:

ACE is built on a composable pipeline engine. Each step declares what it requires and what it produces:

AgentStep -> EvaluateStep -> ReflectStep -> UpdateStep -> DeduplicateStep

Use learning_tail() for the standard learning sequence, or compose custom pipelines:

from ace import Pipeline, AgentStep, EvaluateStep, learning_tail

steps = [AgentStep(agent, skillbook), EvaluateStep(env)] + learning_tail(reflector, skill_manager, skillbook)
pipeline = Pipeline(steps)

The pipeline engine (pipeline/) is framework-agnostic with requires/provides contracts, immutable context, and error isolation. See Pipeline Design and Architecture.

• Full Documentation — Guides, API reference, examples
• Quick Start — 5-minute setup
• Setup Guide — Configuration and providers
• Hosted API Guide — Hosted CLI, trace upload, prompt install
• Architecture — Core concepts and system design
• Code Reference — Implementations, API, usage examples
• Design Decisions — Rejected alternatives and rationale
• Pipeline Engine — Step composition and context flow
• Examples — Runnable demos
• Changelog — Version history

Contributions are welcome. See Contributing Guidelines.