Agentic Context Engine (ACE)

---

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.

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

---

How It Works

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.

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.

---

Runners

Runner	Class	Description
LiteLLM	ACELiteLLM	Batteries-included agent with .ask(), .learn(), .save() — accepts any LiteLLM model string
Core	ACE	Full learning loop with batch epochs and evaluation
Trace Analyser	TraceAnalyser	Learn from pre-recorded traces without re-running tasks
browser-use	BrowserUse	Browser automation that improves with each run
LangChain	LangChain	Wrap any LangChain chain or agent with learning
Claude Code	ClaudeCode	Claude Code CLI tasks with learning

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

---

Benchmarks

Tau2 — Multi-Step Agentic Tasks

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:

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.

---

CLI

Command	Description
ace setup	Interactive setup — model selection, API keys, connection validation
ace models <query>	Search available models with pricing
ace validate <model>	Test a model connection
ace config	Show current configuration
kayba	Cloud CLI — upload traces, fetch insights, manage prompts
ace-mcp	MCP server for IDE integration

---

Documentation

• 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

---

Contributing

Contributions are welcome. See Contributing Guidelines.

---