Know What Your AI Agents Did. Prove It.
Automatic logging of every agent action, decision, and human override. Full EU AI Act Article 12 compliance—without changing how you build agents.
AI Agents Are a Compliance Nightmare
AI agents are transforming what's possible—but from a compliance perspective, they're the highest-risk AI you can deploy.
Autonomous Decision-Making
Unlike simple AI models that respond to prompts, agents decide what to do next. They choose which tools to call, what data to access, and how to complete tasks. Every decision is a potential compliance event.
Multi-Step Complexity
A single agent task might involve dozens of reasoning steps, tool calls, and intermediate decisions. Manual logging is impossible. Missing even one step breaks your audit trail.
Human Oversight Requirements
Article 14 of the EU AI Act requires human oversight for high-risk AI. For agents, this means logging when humans approve, reject, or override agent actions.
Article 12 Is Explicit
The EU AI Act Article 12 specifically requires "automatic recording of events" for the AI system's lifetime. For agents, this means every action, every decision, every intervention—logged immutably.
Existing observability tools track agent performance. Protectron tracks agent compliance. That's the difference between debugging and defending your AI in front of regulators.
Compliance-First Agent Logging
Protectron's Agent Audit Trail captures everything regulators need to see—automatically. One SDK integration, and every agent action is logged with the context required for EU AI Act compliance.
Tool Calls & Results
Every tool your agent invokes is logged with tool name, parameters, execution timestamp, result or error, and duration.
LLM Interactions
Every prompt and completion is captured: input messages, model used, output generated, token usage, and content filtering flags.
Decision Points
When your agent chooses between options: available options considered, selection made, confidence scores, and context that influenced the decision.
Human Oversight Events
Critical for Article 14: human-in-the-loop approvals, rejections with reason, manual overrides, escalation triggers, and intervention timestamps.
Risk & Safety Events
Automatic flagging: content policy violations, unexpected behaviors, error conditions, boundary violations, and anomaly detection triggers.
Session & Context Metadata
Full context: session ID, user context (anonymized), system version, environment, and deployment identifier.
Integration in Minutes, Compliance Forever
Install the SDK
Add Protectron to your project with a single command: pip install protectron or npm install @protectron/sdk
Add the Callback
Integrate with your existing agent framework. One line of code—your business logic stays unchanged.
Run Your Agent
That's it. Run your agent normally. Every action is automatically logged to your Protectron audit trail.
View Audit Trail
Open your Protectron dashboard to see the complete audit trail: timeline of all events, filterable and searchable.
Generate Compliance Evidence
With one click, generate compliance reports that map your audit trail to EU AI Act requirements.
Works With Your Agent Framework
LangChain / LangGraph
The most popular agent framework, fully supported. Works with Agents, Chains, LCEL pipelines, and LangGraph state machines.
from protectron.langchain import ProtectronCallback
agent = create_react_agent(
llm, tools,
callbacks=[ProtectronCallback()]
)CrewAI
Multi-agent orchestration with full crew visibility. Track which agent did what within multi-agent crews.
from protectron.crewai import ProtectronCallback
crew = Crew(
agents=[agent1, agent2],
callbacks=[ProtectronCallback()]
)AutoGen
Microsoft's multi-agent conversation framework. Support for AssistantAgent, UserProxyAgent, and GroupChat.
from protectron.autogen import ProtectronCallback
assistant = AssistantAgent(
"assistant",
callbacks=[ProtectronCallback()]
)Vercel AI SDK
For TypeScript/JavaScript agent builders. Works with streamText, generateText, and tool calls.
import { protectronMiddleware } from '@protectron/vercel-ai';
const result = await generateText({
model: openai('gpt-5.2'),
experimental_telemetry: protectronMiddleware()
});Built for Compliance, Not Just Observability
Human Oversight Tracking
Prove your Article 14 compliance. Log all human approvals, rejections, and overrides with full context and timestamps.
Immutable Audit Trail
Write-once storage with cryptographic verification. Once an event is logged, it cannot be modified or deleted.
Automatic Risk Flagging
Proactive compliance: automatically flag content policy events, boundary violations, anomalies, and custom risk rules.
Compliance Evidence Export
Generate Article 12 and Article 14 compliance reports with one click. PDF format, structured for regulatory review.
Privacy-First Design
PII redaction, content exclusion, EU data residency. Control what gets logged with fine-grained configuration.
7-Year Retention
Default retention aligned with regulatory requirements. Custom retention periods available on Growth and Enterprise plans.
Who Uses Agent Audit Trail
Financial Services Agents
Complete log of every customer interaction. Human approval tracking for sensitive actions. Risk flagging for unusual transaction patterns.
Healthcare AI Assistants
Audit trail for medical-adjacent decisions. Human oversight proof for care recommendations. PII redaction for privacy compliance.
E-commerce Customer Service
Log every customer service decision. Track refund approvals and escalations. Generate evidence for consumer complaints.
Enterprise Workflow Automation
Track agent actions affecting employees. Log approval chains for sensitive operations. Integration with enterprise compliance systems.
Frequently Asked Questions
Logging is asynchronous and non-blocking. Typical latency overhead is <5ms per event. For latency-critical applications, we support batched async logging that adds zero latency to the hot path.
We use a resilient architecture with local buffering. If our servers are unreachable, events are queued locally and synced when connectivity resumes. No events are lost.
Absolutely. Protectron complements observability tools. Use LangSmith for debugging and development, Protectron for compliance. Both callbacks can run simultaneously.
Default retention is 7 years (aligned with typical regulatory requirements). Custom retention periods are available on Growth and Enterprise plans. You can export data before deletion.
Yes. We use append-only storage with cryptographic chaining. Once an event is written, it cannot be modified. Even we cannot alter historical records.
Use our PII redaction features, content exclusion options, or sampling to control what's logged. You can log the fact that an action happened without logging the content of that action.
Make Your Agents Audit-Ready
The EU AI Act treats autonomous AI agents as high-risk. Prove your compliance with automatic audit trails that capture every action, decision, and human intervention.