Skip to Content
Getting StartedCore Concepts

Core Concepts

Kaplaix’s data model is built around four primitives: tenants, agents, sessions, and events.

Tenants

A tenant is an isolated workspace. Every resource — agents, sessions, events, policies — belongs to exactly one tenant. Data is never shared across tenants.

Each tenant has one or more API keys (prefixed al_live_ or al_test_) used to authenticate SDK and HTTP requests.

Agents

An agent is a registered AI agent within a tenant. Each agent has a unique agentId (your chosen identifier, e.g. "support-bot") and metadata like name, framework, and environment.

Agents are created via the API or auto-registered on first event ingestion. Kaplaix builds a behavioral fingerprint (Agent DNA) for each agent over time, detecting drift from established baselines.

Sessions

A session groups a sequence of events from a single agent execution. Think of it as one “run” — from the agent receiving a task to completing it.

  • Sessions are created implicitly when the first event arrives with a new sessionId.
  • Each session tracks eventCount, firstEventAt, lastEventAt, and is resolved to an agent.
  • Sessions can be queried, filtered (by agent, date range, search), and exported as evidence packages.

Events

Events are the atomic unit of observability. Every action your agent takes — tool calls, data access, approvals, environment changes — is captured as an event.

Event envelope

Every event shares a common envelope:

FieldTypeDescription
eventIdUUIDUnique identifier (client-generated)
agentIdstringAgent that produced the event
sessionIdstringSession this event belongs to
sourceTimestampISO 8601When the event occurred
categorystringOne of the 7 canonical categories
schemaVersionstringSchema version (currently "1.0")
payloadobjectCategory-specific data
previousHashstring?Hash of the previous event for chain integrity

The 7 event categories

CategoryPurposeExample
identityAgent identity assertionAgent startup, version declaration
reasoningDecision trace”Decided to escalate to human”
tool_apiTool/API invocationHTTP call, database query, file write
browser_desktopUI interactionScreenshot, click, navigation
data_movementData read/write/exportFile export, record deletion
approvalHuman decisionApprove or reject an action
environmentRuntime contextSandbox mode, network segment

See the full Event Schema reference for payload details.

Risk Engine

The Risk Engine scores every session across 10 dimensions aligned to the OWASP Agentic Top 10  (ASI01–ASI10). Each dimension produces a 0–100 score with explainable factors and actionable recommendations.

Risk is computed at read time from the session’s events — there is no pre-computation. Severity levels: low, medium, high, critical.

Policies

Policies define rules that govern what agents can do. Each policy contains:

  • Rules — conditions matching event attributes (tool type, operation, target system, initiator type)
  • Decisionsallowed, requires_review, or denied
  • Default decision — applied when no rule matches

Only one policy can be active per tenant at a time.

Guardian Mode

Guardian Mode enables pre-execution policy enforcement. Instead of logging events after the fact, the SDK checks with Kaplaix before the dangerous tool runs.

Three enforcement modes:

ModeBehavior
log_onlyLog policy decision, never block
blockBlock denied actions, throw error
require_approvalCreate pending action, wait for human decision

Guardian Mode includes a circuit breaker to protect against cascading failures. See the Guardian API reference for details.

Evidence Packages

Evidence packages are signed ZIP exports containing:

  • PDF report — human-readable session narrative
  • JSON data — machine-readable event stream
  • Manifest — file listing with SHA-256 checksums
  • Digital signature — integrity verification

These are designed for auditors, regulators, and compliance teams.

Last updated on