[!TIP]
ACE is the open-source engine behind Kayba. If you'd rather have the whole loop managed for you, from failure investigation to fixes shipped as PRs, get a demo.
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.
Proven Results
Metric
Result
Context
2x consistency
Doubles pass^4 on Tau2 airline benchmark
15 learned strategies, no reward signals
49% token reduction
Browser automation costs cut nearly in half
10-run learning curve
$1.50 learning cost
Claude Code translated 14k lines to TypeScript
Zero build errors, all tests passing
Quick Start
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.
ACE maintains a Skillbook — a persistent collection of strategies that evolves with every task. Three specialized roles manage the learning loop:
Role
Responsibility
Agent
Executes tasks, enhanced with Skillbook strategies
Reflector
Analyzes execution traces to extract what worked and what failed
SkillManager
Curates the Skillbook — adds, refines, and removes strategies
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.
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.
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.
Claude Code — Autonomous Translation
ACE + Claude Code translated this library from Python to TypeScript with zero supervision:
Metric
Result
Duration
~4 hours
Commits
119
Lines written
~14,000
Build errors
0
Tests
All passing
Learning cost
~$1.50
Pipeline Architecture
ACE is built on a composable pipeline engine. Each step declares what it requires and what it produces:
The pipeline engine (pipeline/) is framework-agnostic with requires/provides contracts, immutable context, and error isolation. See Pipeline Design and Architecture.
CLI
Command
Description
ace setup
Interactive setup — model selection, API keys, connection validation
