ARK — Agent Reliability Kit 🛡

Trust infrastructure for AI agents. Stripe-level idempotency × Sentinel circuit breakers × OpenTelemetry tracing × IDE-style validation. For AI agents.

README.md

💝 赞助 ARK: GitHub Sponsors · 爱发电 · 微信: 84911541@qq.com

ARK 是 MIT 协议开源的。你的赞助让它活下去。详见 SPONSOR.md。

🤔 Why ARK?

Your AI agent says it sent an email. Did it really? Your AI agent says it charged $10. Did it charge $10... or $100? Your AI agent retries on failure. Did it just send 3 duplicate payments?

"Agent does not actually invoke tools, only simulates tool usage with fabricated output" — CrewAI Bug #?, 63 comments

ARK proves every agent action is real.

⚡ Quick Start

pip install ark-trust

from ark import IdempotencyGuard, CircuitBreaker, OutputValidator

# 🛡 1. Never run duplicate payments
guard = IdempotencyGuard()

@guard.wrap
def charge(amount: float):
    return stripe.charge(amount)

charge(99.99)  # ✅ Charged
charge(99.99)  # 🛡 Intercepted — no duplicate!

# ⚡ 2. Auto-fallback when models fail
breaker = CircuitBreaker("gpt-4", failure_threshold=3)

result = breaker.call(
    primary=lambda: gpt4.generate(prompt),
    fallback=lambda: claude.generate(prompt)  # Auto-switch!
)

# 🔧 3. Validate agent output
from pydantic import BaseModel

class PaymentResult(BaseModel):
    amount: float
    txn_id: str

validator = OutputValidator()
result = validator.validate(PaymentResult, agent_output)
if not result.valid:
    print(f"ARK blocked invalid output: {result.errors}")

🔱 The Four Pillars

📊 What ARK Catches

CrewAI agents fake tool calls     → ARK validates real execution
GPT-4 fails 6 times in a row      → ARK auto-switches to Claude
Agent retries a payment 3 times   → ARK blocks duplicates 2&3
Streaming response returns null   → ARK detects + reloads

🎯 The Numbers

• 8847 error issues across top 3 agent frameworks
• 50+ comments on a single "duplicate payment" bug
• 75% of agent outputs need validation
• 0 existing products in this space (max competitor: 129⭐)

🏗 Architecture

Your Agent
    ↓
┌─── ARK Trust Layer ───┐
│  🛡 Guard   ⚡ Breaker │
│  🔧 Validator 👁 Trace │
└───────────────────────┘
    ↓
Your Tools & APIs

🚀 Roadmap

• MVP — 4 core pillars (v0.1.0)
• LangChain integration — ARKCallbackHandler (v0.1.0)
• CrewAI integration — ARKCrewCallback (v0.1.1)
• Auto-detect frameworks (v0.2.0)
• Reliability Score + Schema Registry — 13 schemas (v0.2.0)
• Dashboard UI — real-time trust monitoring (v0.3.0)
• Achievement system — gamified reliability badges (v0.3.0)
• PyPI publish automation (v0.3.0)
• Community Schema Hub (v0.4.0)
• Benchmarks — 7 performance baselines (v0.4.0)
• OpenTelemetry Exporter — 8 reliability event types (v0.5.0)
• Zero-touch instrumentation — one env var to activate (v0.5.0)

🔭 OpenTelemetry Integration (v0.5.0)

ARK reliability events are emitted to your observability stack via standard OTLP/JSON. Zero code changes to existing agents — set one env var:

export ARK_OTEL_ENDPOINT="http://otel-collector:4318/v1/events"

Compatible with Langfuse, Jaeger, Tempo, Honeycomb — any OTLP receiver.

⚡ 5-Minute Live Demo

See ARK events flow into Langfuse with one command:

cd examples/langfuse-demo
docker compose up -d
pip install -e ../..
python app.py
# Open http://localhost:3000 → watch ARK reliability events stream in

🩹 Self-Healing Errors (v0.5.1) — 12-Factor Agents Factor 9

Your agent hits a flaky API. What does it do?

• ❌ Without F9: stack trace explodes the LLM's context window → token waste + confused LLM
• ✅ With ARK F9: error compressed to 500 chars + last 3 stack lines + md5 hash → LLM knows exactly what to fix

from ark.errors import with_retry, should_retry, truncate_error, error_to_llm_context

# 🩹 1. Auto-retry with exponential backoff (1s → 2s → 4s)
@with_retry(tool_name="send_email", max_attempts=3)
def send_email(to, subject):
    return smtp.send(to, subject)

# 🎯 2. Smart retry decisions (8 NON_RETRYABLE_TYPES skipped immediately)
try:
    charge_card(amount)
except AuthError as e:
    if not should_retry(e, attempt=1, max_attempts=3)[0]:
        return redirect_to_human_review()  # Skip retry, save 30s

# 🧠 3. Feed structured error to LLM (self-healing)
try:
    call_external_api()
except Exception as e:
    prompt = error_to_llm_context(e)  # 500-char message + stack tail + retry hint
    response = llm.invoke(prompt)     # LLM decides: different tool? different args?

⚡ 1-Minute F9 Demo (no Docker)

cd examples/f9-self-healing
python app.py
# See: 5KB stack compressed to ~500 chars, 8 NON_RETRYABLE_TYPES identified, 3-round retry simulation

🌉 Native OTel SDK Bridge (v0.5.3) — Zero-friction observability

One-line upgrade for OTel users. If your stack already runs the OpenTelemetry SDK, ARK now dual-emits reliability events to your existing tracer — no code change, no OTLP collector required.

# Before v0.5.3: OTLP/JSON only (need collector)
export ARK_OTEL_ENDPOINT=http://collector:4318

# After v0.5.3: dual-emit, automatic if opentelemetry-api installed
pip install opentelemetry-api   # that's it
export ARK_OTEL_ENDPOINT=http://collector:4318  # collector still works
# Native spans flow into your existing tracer (Jaeger/Tempo/Honeycomb) automatically

Why this matters:

• 🔌 Plug into existing observability — If you already pay for Datadog/Honeycomb/Tempo, ARK events appear next to your app spans
• 🛡 100% backward compatible — use_native_sdk=False (default) keeps the old OTLP/JSON path; opt in with one flag
• 🪫 Zero overhead when off — if not use_native_sdk: return is the first line; no imports, no allocations
• 🚨 Failure-isolated — try/except around native span emit: a broken SDK init never breaks your OTLP export
• 🟥 Auto-set ERROR status on VALIDATION_FAIL — your alerting tools see it immediately

Test coverage: 16 new tests in test_v0_5_3_otel_sdk_bridge.py covering type coercion, SDK availability probing, opt-out path, span emission, attribute propagation, ERROR status on validation failure, failure isolation, stats() reporting, and backward-compat signatures.

📦 What's New in v0.5.3

📜 License

MIT — Free forever. ARK is open infrastructure.

Built with 🧬 gene recombination: Stripe × Sentinel × OpenTelemetry × IDE