bx-ai: AI Agents

BoxLang is the AI-native software productivity platform for building, modernizing and running applications, with developers and AI agents working together.

aiAgent() BIF

// Signature (all optional)
aiAgent(
    name            = "BxAi",
    description     = "",
    instructions    = "",
    model           = aiModel(),      // aiModel() instance
    memory          = aiMemory(),     // one IAiMemory or an array of them
    tools           = [],             // array of aiTool() instances
    subAgents       = [],             // other agents, auto-wrapped as tools
    skills          = [],             // always-on AiSkill instances
    availableSkills = [],             // lazy-loaded skills pool
    params          = {},             // default model params
    options         = {},             // default run options (threadId, userId, conversationId, ...)
    middleware      = [],             // IAiMiddleware, struct of closures, or array of either
    checkpointer    ,                 // IAiMemory used to save state when a run suspends
    checkpointTTL   = 30,             // checkpoint TTL in minutes
    mcpServers      = [],             // MCP server URLs / HTTP or stdio config structs / IMCPClient instances
    register        = false,          // register in aiAgentRegistry()
    module          = ""              // registry key suffix when register is true
)

Basic Agent

agent = aiAgent(
    name        : "Assistant",
    description : "A helpful assistant",
    instructions: "Be concise, accurate, and friendly."
)

response = agent.run( "What is BoxLang?" )
println( response )

run( input, params = {}, options = {} ) returns the assistant content (returnFormat defaults to single).

Agent with a Custom Model

model = aiModel( provider: "claude", params: { model: "claude-sonnet-5" } )

agent = aiAgent(
    name  : "Claude Agent",
    model : model,
    params: { temperature: 0.7, max_tokens: 2000 }
)

Agent with Tools

weatherTool = aiTool(
    "get_weather",
    "Get current weather for a location",
    location -> getWeatherData( location )
).describeLocation( "City name, e.g. Boston, MA" )

agent = aiAgent(
    name        : "DataAgent",
    instructions: "Use tools to fetch real-time data. Always be accurate.",
    tools       : [ weatherTool ]
)

response = agent.run( "What's the weather in Boston?" )

aiTool( name, description, callable, autoRegister = true ). Fluent describers: .describeMyArg( "..." ), .describeArg( name, text ), .describe( text ). See the bx-ai-tools skill for more.

Agent with Skills

Skills inject markdown knowledge into the system message.

agent = aiAgent(
    name           : "CodeReviewer",
    instructions   : "Review code for quality and security issues",
    // Always-on: full content injected on every call (a file path returns ONE AiSkill)
    skills         : [
        aiSkill( ".agents/skills/security/SKILL.md" ),
        aiSkill( ".agents/skills/code-style/SKILL.md" )
    ],
    // Lazy: only an index is injected, the LLM loads one via the built-in loadSkill tool
    // (a directory path returns an ARRAY of AiSkill, scanned recursively)
    availableSkills: aiSkill( ".agents/skills/languages" )
)

aiSkill( path, name, description = "", content = "", recurse = true ). With path it loads from disk; with name (and no path) it builds an inline skill:

aiAgent(
    name  : "Writer",
    skills: [
        aiSkill(
            name       : "tone",
            description: "Professional writing tone",
            content    : "Always write in a clear, professional tone. Avoid jargon."
        )
    ]
)

Agent with Memory

// "window" keeps the last N messages (default 100)
memory = aiMemory( "window", config: { maxMessages: 20 } )

agent = aiAgent( name: "ConversationalBot", memory: memory )

agent.run( "My name is Alice." )
agent.run( "What is my name?" )  // "Your name is Alice."

aiMemory( memory, key, userId, conversationId, config ). Common types: window, cache, file, jdbc, summary, session, hybrid, plus vector stores. Shared memory is scoped per call with options: { userId, conversationId }. See bx-ai-memory.

Agent with MCP Servers

Each entry is a URL string, an HTTP config struct (url), a stdio config struct (command), or an IMCPClient built with MCP(). The tools the server lists are loaded, all of them unless toolFilter is set. See bx-ai-mcp for the full MCP client guide.

agent = aiAgent(
    name      : "ResearchAgent",
    mcpServers: [
        "http://localhost:3000/mcp",
        { url: "https://api.example.com/mcp", token: server.system.environment.MCP_TOKEN, timeout: 5000 },
        // A program the agent launches as a child process (stdin/stdout)
        { command: [ "bxplaywright", "mcp" ], env: { HEADLESS: "true" }, prefix: "browser" }
    ]
)

agent.run( "Open boxlang.io and read the title" )

// Shut the stdio child processes down (clients the agent built itself)
agent.closeMCPServers()

HTTP struct keys: url, token (bearer), timeout, headers, user + password (basic auth), urlGuard, redirects, maxResponseBytes, handshake, httpVersion. Stdio struct keys: command (array, required), env, cwd, inheritEnv, maxRestarts, commandGuard, timeout. Both: prefix (tools become prefix__name) and toolFilter (only these tools). Also available as withMCPServer( server, config ) and withMCPServers( servers ).

Fluent Configuration

agent = aiAgent( name: "Assistant" )
    .setModel( aiModel( provider: "claude" ) )
    .setInstructions( "Be concise and helpful" )
    .addTool( searchTool )
    .addMemory( conversationMemory )
    .setParam( "temperature", 0.6 )
    .withMiddleware( new bxModules.bxai.models.middleware.core.LoggingMiddleware() )
    .withCheckpointer( aiMemory( "cache" ), 30 )

Other setters: setName, setMemory (replaces), setTools, withTools, addSubAgent, setSubAgents.

Streaming Agent Responses

The callback comes FIRST: stream( onChunk, input, params = {}, options = {} ).

agent.stream(
    chunk -> print( chunk.choices.first().delta?.content ?: "" ),
    "Explain how Hibernate ORM works"
)

Chunks are OpenAI-shaped structs. Reasoning-capable models also surface choices[].delta.reasoning (streaming) and choices[].message.reasoning (sync run), kept separate from content and never saved to memory. See bx-ai-chatting for detail.

Sub-Agents and Multi-Agent Hierarchy

subAgents are registered as tools named delegate_to_<agent-name-slug>, described from each sub-agent's name and description. Give sub-agents a clear description.

coder    = aiAgent( name: "Coder",    description: "Writes clean BoxLang code", instructions: "Write clean BoxLang code" )
tester   = aiAgent( name: "Tester",   description: "Writes TestBox tests",      instructions: "Write TestBox test cases" )
reviewer = aiAgent( name: "Reviewer", description: "Reviews code for bugs",     instructions: "Review code for bugs and performance" )

router = aiAgent(
    name        : "Router",
    instructions: "Route tasks to the appropriate specialist agent",
    subAgents   : [ coder, tester, reviewer ]
)

result = router.run( "Write a BoxLang class for user authentication and tests for it" )

Inspect with getSubAgentNames(), getSubAgent( name ), hasSubAgent( name ), getAgentPath().

Middleware

Pass one or many via middleware: (or .withMiddleware()). Built-ins live in bxModules.bxai.models.middleware.core.

import bxModules.bxai.models.middleware.core.LoggingMiddleware;
import bxModules.bxai.models.middleware.core.RetryMiddleware;
import bxModules.bxai.models.middleware.core.MaxToolCallsMiddleware;
import bxModules.bxai.models.middleware.core.GuardrailMiddleware;

agent = aiAgent(
    name      : "SafeAgent",
    tools     : [ searchTool, sqlTool ],
    middleware: [
        new LoggingMiddleware( logToConsole: true ),
        new RetryMiddleware( maxRetries: 3 ),
        new MaxToolCallsMiddleware( maxCalls: 5 ),
        new GuardrailMiddleware(
            blockedTools: [ "dropTable" ],
            argPatterns : { runSql: [ { query: "(?i)drop|truncate|delete" } ] }
        )
    ]
)

Record then replay an agent offline (zero live calls):

aiAgent( middleware: new FlightRecorderMiddleware( mode: "record" ) ).run( "Weather in London?" )
aiAgent( middleware: new FlightRecorderMiddleware( mode: "replay", fixturePath: "tests/fixtures/london.json" ) )

Security middleware (InputSanitizerMiddleware, OutputGuardMiddleware, LLMGuardMiddleware) lives in models.middleware.security.

Human in the Loop (HITL)

HumanInTheLoopMiddleware pauses a tool call until a human decides.

init( toolsRequiringApproval = [], mode = "cli", showArguments = true,
      approvalCallback, policy, gateway, decisionStore )

Any middleware that can suspend (a gateway other than cli, or mode: "web") REQUIRES a checkpointer: on the agent, or attaching it throws HumanInTheLoopMiddleware.NoCheckpointer.

import bxModules.bxai.models.middleware.core.HumanInTheLoopMiddleware;

agent = aiAgent(
    name        : "OrderAgent",
    tools       : [ placeOrderTool ],
    middleware  : new HumanInTheLoopMiddleware( toolsRequiringApproval: [ "placeOrder" ], mode: "web" ),
    checkpointer: aiMemory( "cache" )
)

result = agent.run( "Order 3 widgets", {}, { threadId: "order-42" } )

if ( isObject( result ) && result.isSuspended() ) {
    // show result.getData() to a human, then later:
    answer = agent.resume( "approve", "order-42" )
    // or: agent.resume( "reject", "order-42", {}, "alice", "too expensive" )
    // or: agent.resume( "edit", "order-42", { /* corrected data */ } )
}

resume( decision, threadId, editedData = {}, decidedBy = "", reason = "" ) and the streaming mirror resumeStream( onChunk, decision, threadId, editedData, decidedBy, reason ). resume throws AiAgent.NoCheckpointer or AiAgent.CheckpointNotFound when it cannot continue. The decision strings are approve, reject, edit, cancel, approve_always, approve_session.

Batched approvals: when the model requests several tool calls at once, decision may be one string applied to all pending calls, or an array of per-call structs { decision, editedData?, decidedBy?, reason? } in the original order. Calls that did not need approval are not re-run.

Policies decide which calls need approval (pass policy:; default is by tool name). In models.hitl.policies:

Durable grants: approve_always is stored in an IDecisionStore; approve_session is in-memory per thread and tool. Create one with aiDecisionStore( store, config ) where store is cache, jdbc, or file, and pass it as decisionStore:. Omitted, it uses settings.hitl.decisionStore (default cache).

new HumanInTheLoopMiddleware(
    policy       : new RiskLevelApprovalPolicy( minLevel: "high" ),
    gateway      : aiGateway( "mock" ),
    decisionStore: aiDecisionStore( "file", { directoryPath: "/data/grants" } )
)

Run Control

Every run has a threadId: pass options: { threadId } to run()/stream(), otherwise a UUID is generated. From anywhere (a webhook, an admin endpoint) address the in-flight run by that id. Both return true if a run was in flight and false otherwise, and take effect at the run's next LLM or tool checkpoint, not mid-token.

future = asyncRun( () -> agent.run( "Long research task", {}, { threadId: "job-7" } ) )

agent.steerRun( "job-7", "Focus on 2025 data only" )       // splice a user message into the running turn
agent.cancelRun( "job-7", "User pressed stop" )             // reason is optional

Gateways

A gateway connects an agent to a channel (CLI, HTTP, chat platforms). Core gateways: cli, http, mock (for tests). Other names resolve through aiGatewayRegistry() (e.g. external gateway modules).

gateways takes instances or names ([ "cli" ]). The policy decides what happens when a message arrives on a thread that already has a turn running:

agent = aiAgent(
    name        : "Support",
    checkpointer: aiMemory( "cache" ),
    middleware  : new HumanInTheLoopMiddleware( toolsRequiringApproval: [ "refund" ], gateway: aiGateway( "cli" ) )
)

aiGatewaySession( agent: agent, gateways: [ "cli" ], policy: "steer" ).start()

HTTP front controller (models.gateway.http.GatewayRequestProcessor, all static, each returns { statusCode, body, contentType, headers }). The HTTP gateway must be registered first: aiGatewayRegistry().register( aiGateway( "http", { secret: "..." } ) ).

Pass the exact raw body string, since signatures are computed over it.

Untrusted Content: aiFence()

aiFence( content, label = "external", withPreamble = false ) wraps untrusted text (RAG documents, tool or web output) in boundary markers so the model treats it as data, not instructions.

context = aiFence( retrievedDoc, "knowledge-base", withPreamble: true )
agent.run( "Answer using this context: #context#" )

Common Pitfalls

• Do NOT call .build() on aiAgent(), it doesn't exist
• Do NOT call .withMemory(): pass memory: in the constructor or use .addMemory() / .setMemory()
• aiMemory( "windowed" ) throws InvalidMemoryType: the type is "window"
• agent.stream() takes the callback first, then the input
• mcpServers entries have no toolNames or apiKey key: use token
• A stdio entry takes command (an array), not url. HTTP-only keys (token, headers, urlGuard, ...) on a stdio entry throw InvalidArgument
• Call agent.closeMCPServers() when done with stdio servers, otherwise the child processes live until the module unloads
• A stdio server that cannot start is logged and skipped (the agent runs without its tools); a command refused by mcp.stdio.allowedCommands or commandGuard throws MCPCommandRefused
• A suspended run() returns an AiMiddlewareResult, not a string: check isSuspended()
• A HITL gateway or mode: "web" without an agent checkpointer throws at attach time
• Always give agents clear, specific instructions and sub-agents a clear description
• Use availableSkills (lazy) for large skill libraries; use skills (always-on) only for core rules