Lite SuiteLite Suite
← Back to Docs

LiteAgent

Self-improving autonomous AI agent that manufactures its own tools — runs 100% local

Overview

LiteAgent is a self-improving autonomous AI agent built into Lite Suite. It has identity files, heartbeat scheduling, a Python CLI agent loop, DAG-based memory, and a self-improvement loop that auto-creates skills from completed sessions — optionally compiling them into executable MCP tools via LiteCLI.

The entire system runs locally on a 24B parameter model (Devstral via LM Studio). Zero API cost. No data leaves your machine.

What Makes It Different

Every other AI agent that "learns" writes prose — markdown instructions injected into prompts. LiteAgent can compile skills into real executable tools via LiteCLI. The agent doesn't just remember how to do something — it manufactures a new tool that does it.

| Feature | Other Agents | LiteAgent | |---------|-------------|-----------| | Learning | Prompt injection (prose) | Skills + compiled executable tools | | LLM required | Frontier APIs | Local 24B model ($0) | | Data privacy | Sent to API provider | 100% local | | Skill format | Proprietary | agentskills.io compatible (open standard) |

Self-Improvement Loop

After completing a task, a background reviewer analyzes the session:

  1. Detects trial-and-error approaches, multi-step workflows, user corrections, and domain knowledge
  2. Creates skills as SKILL.md files (agentskills.io format) with security scanning
  3. Updates identity — appends to memory.md and user.md
  4. Compiles API-wrapping skills into executable CLI tools via LiteCLI (optional)

Skills auto-activate immediately. The reviewer LLM is configurable.

Three Trigger Points

  • Automatic — background review after every completed task
  • Heartbeat sweep — periodic consolidation of unreviewed sessions
  • Manual — /review, /review all, /review <session-id> commands

LiteCLI Compilation

When auto_compile is enabled, skills that wrap HTTP APIs get compiled into standalone Python CLI tools:

SKILL.md → LiteCLI compiler → executable .py CLI → registered as MCP tool

Next time the agent runs, it discovers the new tool via MCP and can call it directly — not as a prompt instruction, but as an actual executable tool.

Identity System

Four core files give the agent a persistent sense of self:

| File | Purpose | |------|---------| | soul.md | Personality, values, and tone | | user.md | Information about who the agent serves | | memory.md | Long-term memories (updated by reviewer) | | heartbeat.md | Instructions for periodic wake-ups |

Python CLI

liteagent agent              # Interactive REPL
liteagent agent -p "..."     # One-shot prompt
liteagent agent --restricted # Read-only scout mode
liteagent review             # Review most recent session
liteagent review --all       # Sweep all unreviewed sessions

Progressive Skill Loading

4-tier system that minimizes token usage:

| Tier | What Loads | Cost | |------|-----------|------| | 0 | Category names only | Cheapest | | 1 | Name + description | Low | | 2 | Full SKILL.md content | Medium | | 3 | Supporting files | On demand |

Configuration

Reviewer config at ~/.litesuite/agent/config/reviewer-config.json:

{
  "enabled": true,
  "model": "",
  "base_url": "",
  "auto_compile": false,
  "max_review_tokens": 4000,
  "min_session_messages": 4
}

Leave model and base_url empty to use the same LM Studio model the agent uses for tasks.

Heartbeat Service

A reserved scheduled job. On each tick:

  1. Assembles identity context from soul/user/memory/heartbeat files
  2. Injects LCM DAG context (summaries + recent messages)
  3. Spawns claude CLI or liteagent agent (LM Studio)
  4. After completion, runs review sweep on unreviewed sessions
  5. Created skills and identity updates broadcast to the Lite Suite UI