AI Agents Overview

Vision

Logrite’s agentic AI system enables users to create autonomous AI agents that monitor logs, detect anomalies, generate reports, and take action — all without requiring manual intervention.

Agents are:

• Created via natural language — “Monitor payment errors and alert me if the error rate exceeds 5%”
• Executed on schedules — Cron-based or event-triggered
• Governed by permissions — Each agent has an explicit list of tools it can access
• Workspace-isolated — Agents only access data from their owning workspace

Current State vs. Planned State

Feature	Current Status	Phase
AI chat with tool calling	✅ Live	—
NLP-based agent creation	✅ Live (2026-07-19)	Phase 1
Tool governance registry	✅ Live (2026-07-19)	Phase 1
Manual (on-demand) execution	✅ Live (2026-07-19)	Phase 1
Execution history UI	✅ Live (2026-07-19)	Phase 1
Background Agent Runner	🔧 Planned	Phase 2
Scheduled execution	🔧 Planned	Phase 2
Alert conditions	🔧 Planned	Phase 3
Write / self-healing tools	🔧 Planned	Phase 3
Multi-agent orchestration	🔧 Planned	Phase 5

Note

Phase 1 shipped (2026-07-19). AI agents are now a first-class, tool-governed feature distinct from collector agents. Backend: ai_agent + agent_execution_log tables and /api/v1/ai-agents CRUD + execution endpoints. AI-service: /api/v1/agent-builder/parse (NL description → reviewable config), a tool-governance registry (src/tools/governance.ts, read/write/external — Phase 1 enforces read-only server-side), and /api/v1/agent-builder/run (manual run → AgentLoop with the agent’s toolsAllowed → captured execution log). Frontend: the AI Agents page (/ai-agents) with an NL AgentBuilder, a review screen, run-on-demand, and the execution-history timeline (the agent’s answer + the exact tool calls it made). Scheduling and write / self-healing tools are the next phases.

Philosophy

Natural Language First

Creating an agent should feel like talking to a colleague: “Watch for database query timeouts and send me a daily summary.” The system extracts the intent, suggests tools and a schedule, and presents a review screen before creating anything.

Explicit Tool Governance

Agents don’t get unlimited access. Every agent has a toolsAllowed list scoped to the minimum necessary. A monitoring agent doesn’t need write access. An alert agent doesn’t need admin tools.

Transparent Execution

Every agent execution is logged with the full AI response, tools called, and outcome. Users can inspect what the agent did and why.

Human in the Loop

Agents propose; humans approve. NLP creation always goes through a review screen. Alert thresholds are set explicitly. Critical actions (deployments, deletions) require human confirmation.

Quick Navigation

Architecture

Full system diagrams and service interactions

Agent Types

System, user-created, and scheduled agents

NLP Creation

How natural language becomes an agent config

Tool Governance

Tool registry, permission levels, categories

Execution Loop

Background scheduler and AgentRunner flow

Roadmap

5 implementation phases with milestones