Micro AGI

Build a self-modifying agent that creates its own tools at runtime — reading docs, writing Lua, registering entries in the registry, and loading them into the active session.

What We're Building

A terminal agent that:

Answers questions using an LLM with streaming
Searches Wippy documentation to learn APIs
Inspects the registry to discover existing capabilities
Builds new tools on the fly when it lacks a capability
Manages its own context window via compression

flowchart LR
    User -->|prompt| Agent
    Agent -->|step| LLM[GPT-5.1]
    LLM -->|tool_calls| Agent
    Agent -->|funcs.call| Tools
    Tools -->|result| Agent
    Agent -->|text| User

    subgraph Tools
        doc_search
        registry_list
        registry_read
        create_tool
        load_tool
    end

Architecture

The agent runs as a Wippy process with access to the registry. When the LLM decides it needs a capability it doesn't have, it uses the self-modification loop:

sequenceDiagram
    participant U as User
    participant A as Agent
    participant L as LLM
    participant R as Registry

    U->>A: "what time is it?"
    A->>L: step(conversation)
    L->>A: tool_call: doc_search("lua/core/time")
    A->>A: execute doc_search
    A->>L: step(conversation + tool result)
    L->>A: tool_call: create_tool(name, source, schema)
    A->>R: evaluate ns deny policies + changeset create
    R->>A: ok
    A->>L: step(conversation + tool result)
    L->>A: tool_call: load_tool("app.generated:current_time")
    A->>A: ctx:add_tools() + reload agent
    A->>L: step(conversation + tool result)
    L->>A: tool_call: current_time()
    A->>A: execute new tool
    A->>L: step(conversation + tool result)
    L->>A: text: "The current time is..."
    A->>U: stream response

The key insight: tools are registry entries. Creating a tool is just writing a function.lua entry with inline Lua source in data.source. The agent runtime compiles and loads it like any other entry.

Project Structure

micro-agi/
├── .wippy.yaml
├── wippy.yaml
└── src/
    ├── _index.yaml
    ├── README.md
    ├── agent.lua
    └── tools/
        ├── _index.yaml
        ├── doc_search.lua
        ├── registry_list.lua
        ├── registry_read.lua
        ├── create_tool.lua
        └── load_tool.lua

Infrastructure

Create .wippy.yaml:

version: "1.0"

logger:
  mode: development
  level: info
  encoding: console

Entry Definitions

Create src/_index.yaml with infrastructure, security policies, models, agent, and process:

version: "1.0"
namespace: app

entries:
  - name: definition
    kind: ns.definition
    readme: file://README.md
    meta:
      title: Micro AGI
      description: Self-modifying development agent that builds its own tools at runtime
      depends_on: [wippy/llm, wippy/agent]

  - name: os_env
    kind: env.storage.os

  - name: processes
    kind: process.host
    lifecycle:
      auto_start: true

  - name: __dep.llm
    kind: ns.dependency
    component: wippy/llm
    version: "*"
    parameters:
      - name: env_storage
        value: app:os_env
      - name: process_host
        value: app:processes

  - name: __dep.agent
    kind: ns.dependency
    component: wippy/agent
    version: "*"
    parameters:
      - name: process_host
        value: app:processes

Security Policies

Two security.policy entries use namespace-based deny to restrict where the agent can write:

  # deny tool writes to core namespace
  - name: deny_core_ns
    kind: security.policy
    policy:
      actions: "*"
      resources: "app:*"
      effect: deny
    groups:
      - agent_security

  # deny tool writes to tools namespace
  - name: deny_tools_ns
    kind: security.policy
    policy:
      actions: "*"
      resources: "app.tools:*"
      effect: deny
    groups:
      - agent_security

These policies are loaded as a named scope (app:agent_security) by create_tool and evaluated before any registry write. The agent can write to app.generated:* (no deny policy matches), but cannot write to app:* (core entries, models, agent definition) or app.tools:* (built-in tools).

See Security Model for details on policy evaluation.

Models

Two models serve different purposes:

  # reasoning model for the agent
  - name: gpt-5.1
    kind: registry.entry
    meta:
      name: gpt-5.1
      type: llm.model
      title: GPT-5.1
      comment: Reasoning model
      capabilities: [generate, tool_use, structured_output, vision, thinking]
      class: [reasoning]
      priority: 210
    max_tokens: 128000
    output_tokens: 32768
    pricing:
      input: 2.5
      output: 10
    providers:
      - id: wippy.llm.openai:provider
        options:
          reasoning_model_request: true
        provider_model: gpt-5.1
    thinking_effort: 10

  # cheap model for context compression
  - name: gpt-4.1-nano
    kind: registry.entry
    meta:
      name: gpt-4.1-nano
      type: llm.model
      title: GPT-4.1 Nano
      comment: Compression model
      capabilities: [generate, tool_use, structured_output]
      class: [fast]
      priority: 100
    max_tokens: 1047576
    output_tokens: 32768
    pricing:
      input: 0.1
      output: 0.4
    providers:
      - id: wippy.llm.openai:provider
        provider_model: gpt-4.1-nano

GPT-5.1 handles reasoning and tool use. GPT-4.1 Nano handles context compression at 25x lower cost.

Agent Definition

  - name: dev_assistant
    kind: registry.entry
    meta:
      type: agent.gen1
      name: dev_assistant
      title: Dev Assistant
      comment: Wippy development assistant
    prompt: |
      Self-modifying Wippy development agent. You run inside Wippy runtime
      with access to docs, registry, and dynamic tool creation.

      Rules:
      - NEVER fabricate, guess, or hallucinate facts. If you need real data,
        use or build a tool to get it. Only state what a tool actually returned.
      - Maximum 2-3 sentences per response. No bullet lists. No disclaimers.
      - Never say "I can't" or "I don't have". Build the tool and do it.
      - Act first, explain only if asked.

      To gain new capabilities: doc_search the API, create_tool with Lua source,
      load_tool, call it. All in one turn.      
    model: gpt-5.1
    max_tokens: 2048
    tools:
      - "app.tools:*"

The prompt is deliberately terse. Key rules:

No hallucination — the agent must use tools for real data
Self-modification — build tools instead of refusing
Action over explanation — do first, explain if asked

Process

  - name: agent
    kind: process.lua
    meta:
      command:
        name: agent
        short: Start dev assistant
    source: file://agent.lua
    method: main
    modules: [io, json, process, channel, funcs, registry, time, security]
    imports:
      prompt: wippy.llm:prompt
      agent_context: wippy.agent:context
      compress: wippy.llm.util:compress

The process runs as a terminal command (root level). Security enforcement happens inside the tools themselves — create_tool loads the agent_security policy group and evaluates it before writing.

Imports:

prompt — conversation builder
agent_context — agent loading and dynamic tool management
compress — LLM-based text compression for context management

Tools

Create src/tools/_index.yaml with five tools:

doc_search

Fetches Wippy documentation via the wippy.ai/llm API. Supports two modes: fetch a page by path, or search by query.

local http_client = require("http_client")
local json = require("json")

local BASE_URL = "https://wippy.ai/llm"
local MAX_CHARS = 8000

local function fetch_page(path)
    local url = BASE_URL .. "/path/en/" .. path
    local resp, err = http_client.get(url, {
        headers = { ["User-Agent"] = "wippy-agent/1.0" },
    })
    if err then
        return nil, tostring(err)
    end
    if resp.status_code ~= 200 then
        return nil, "HTTP " .. resp.status_code
    end

    local body = resp.body or ""
    if #body > MAX_CHARS then
        body = body:sub(1, MAX_CHARS) .. "\n... (truncated)"
    end
    return body, nil
end

local function search_docs(query)
    local url = BASE_URL .. "/search?q=" .. query
    local resp, err = http_client.get(url, {
        headers = { ["User-Agent"] = "wippy-agent/1.0" },
    })
    if err then
        return { error = tostring(err) }
    end
    if resp.status_code ~= 200 then
        return { error = "HTTP " .. resp.status_code }
    end

    local body = resp.body or ""
    if #body > MAX_CHARS then
        body = body:sub(1, MAX_CHARS) .. "\n... (truncated)"
    end

    return { results = body }
end

local function handler(input)
    if input.path then
        local content, err = fetch_page(input.path)
        if err then
            return { error = err }
        end
        return { path = input.path, content = content }
    end

    if input.query then
        return search_docs(input.query)
    end

    return { error = "provide either 'path' or 'query'" }
end

return { handler = handler }

create_tool

The core of self-modification. Validates input, evaluates namespace deny policies, and creates a function.lua entry in the registry with inline Lua source.

Source validation — code-level checks that policies cannot enforce:

local registry = require("registry")
local json = require("json")
local security = require("security")

local NAMESPACE = "app.generated"
local MAX_SOURCE_LEN = 16000
local MAX_NAME_LEN = 64

local ALLOWED_MODULES = {
    time = true, json = true, http_client = true, expr = true,
    text = true, base64 = true, yaml = true, crypto = true,
    hash = true, uuid = true, url = true,
}

local BLOCKED_PATTERNS = {
    "os%.execute", "os%.remove", "os%.rename",
    "io%.open", "io%.popen",
    "loadfile", "dofile",
    "debug%.", "package%.",
    "rawset", "rawget",
    "setfenv", "getfenv",
}

Policy evaluation — create_tool loads the agent_security named scope and evaluates the deny policies against the target entry ID. Writes to app:* or app.tools:* are denied; writes to app.generated:* pass (no matching deny policy):

local actor = security.new_actor("service:agent", { role = "agent" })
local scope, scope_err = security.named_scope("app:agent_security")
if scope_err then
    return { error = "failed to load security scope: " .. tostring(scope_err) }
end

local result = scope:evaluate(actor, action, id)
if result == "deny" then
    return { error = "policy denied: " .. action .. " on " .. id }
end

Registry write — the entry is written with source in data.source:

local entry = {
    id = id,
    kind = "function.lua",
    meta = {
        type = "tool",
        title = input.name,
        comment = input.description,
        input_schema = schema,
        llm_alias = input.name,
        llm_description = input.description,
    },
    data = {
        source = input.source,
        modules = modules,
        method = "handler",
    },
}

local snap = registry.snapshot()
local changes = snap:changes()
if existing then
    changes:update(entry)
else
    changes:create(entry)
end
changes:apply()

No files on disk. The tool lives entirely in the registry.

load_tool

Validates the entry is a tool and signals the agent loop to reload:

local function handler(input)
    local entry, err = registry.get(input.id)
    if err then
        return { error = tostring(err) }
    end
    if not entry then
        return { error = "not found: " .. input.id }
    end
    if not entry.meta or entry.meta.type ~= "tool" then
        return { error = "not a tool (meta.type != 'tool'): " .. input.id }
    end

    return {
        loaded = true,
        id = entry.id,
        alias = entry.meta.llm_alias or input.id,
        description = entry.meta.llm_description or "",
    }
end

The agent loop detects loaded = true in the result and calls ctx:add_tools(id) followed by ctx:load_agent() to recompile the agent with the new tool.

Agent Loop

The agent loop in src/agent.lua handles streaming, tool execution, dynamic loading, and context compression.

Streaming

Uses the same coroutine + channel pattern from the LLM Agent tutorial:

coroutine.spawn(function()
    local response, err = session.runner:step(session.conversation, {
        stream_target = {
            reply_to = process.pid(),
            topic = STREAM_TOPIC,
        },
    })
    done_ch:send({ response = response, err = err })
end)

Tool Execution

Tools are called via funcs.call() with pcall for safety:

local ok, result = pcall(funcs.call, tc.registry_id, args)

Dynamic Tool Loading

When load_tool returns loaded = true, the agent reloads itself:

flowchart TD
    A[load_tool returns loaded=true] --> B[ctx:add_tools id]
    B --> C[ctx:load_agent]
    C --> D[New runner with added tool]
    D --> E[Conversation preserved]
    E --> F[Next LLM step sees new tool]

local function handle_tool_loading(tool_calls, results)
    local reload_needed = false
    for _, tc in ipairs(tool_calls) do
        if tc.name == "load_tool" then
            local result = results[tc.id]
            if result and result.loaded then
                session.ctx:add_tools(result.id)
                reload_needed = true
            end
        end
    end
    if reload_needed then
        reload_agent()
    end
end

The conversation is preserved across reloads because it lives in the prompt builder, not in the runner.

Context Compression

When prompt tokens exceed 96K (75% of the 128K context window), the conversation is compressed using GPT-4.1 Nano:

if response.tokens and response.tokens.prompt_tokens
    and response.tokens.prompt_tokens > PROMPT_TOKEN_LIMIT then
    try_compress()
end

Compression extracts message content, calls compress.to_size() targeting 4000 characters, and replaces the conversation with a summary:

local summary = compress.to_size(COMPRESS_MODEL, full_text, COMPRESS_TARGET)
session.conversation = prompt.new()
session.conversation:add_system("Conversation summary:\n\n" .. summary)

Security Model

The agent is secured at two levels: declarative namespace deny policies and code-level source validation.

flowchart TD
    LLM[LLM generates tool] --> P{Namespace Deny Policies}
    P -->|scope:evaluate| Check{Target namespace?}
    Check -->|app.generated:*| OK[No deny match - proceed]
    Check -->|app:* or app.tools:*| Deny1[Policy Denied]

    OK --> V{Source Validation}
    V -->|module check| M[allowlist only]
    V -->|pattern check| B[no os.execute, io.open, etc.]
    V -->|size check| S[max 16KB]
    M & B & S -->|all pass| R[Registry write]
    V -->|any fail| Deny2[Validation Error]

Namespace Deny Policies

Policy	Resources	Effect
`deny_core_ns`	`app:*`	deny
`deny_tools_ns`	`app.tools:*`	deny

create_tool loads the agent_security policy group and evaluates against the target entry ID. Since deny policies only match app:* and app.tools:*, writes to app.generated:* pass through (result is undefined, meaning "not denied").

This prevents the agent from:

Modifying its own prompt or agent definition (app:dev_assistant)
Overwriting its built-in tools (app.tools:*)
Changing infrastructure entries (app:processes, etc.)

Source validation additionally blocks os.execute, io.open, debug.*, package.*, and enforces a module allowlist and 16KB size limit.

Run

Run directly from hub:

export OPENAI_API_KEY=sk-...
wippy run wippy/micro-agi agent

Or clone and run locally:

cd micro-agi
export OPENAI_API_KEY=sk-...
wippy init && wippy update
wippy run agent

dev assistant (quit to exit)

> what time is it?
  [doc_search] ok
  [create_tool] ok
  [load_tool] ok
  [+] app.generated:current_time_utc
  [current_time_utc] ok
The current UTC time is 2026-02-13T03:13:41Z.

> fetch https://httpbin.org/get and show my ip
  [create_tool] ok
  [load_tool] ok
  [+] app.generated:http_get
  [http_get] ok
Your IP is 98.24.33.45.

Next Steps

LLM Agent — Build a basic agent from scratch
Agent Module — Agent framework reference
Registry — How the registry works
Security Model — Declarative security policies
Entry Kinds — Available entry types