Remote Tools

AgentEnsemble supports cross-language tool implementations through two remote tool types: ProcessAgentTool (subprocess execution) and HttpAgentTool (REST endpoint wrapping). Both implement AgentTool and are indistinguishable from native Java tools from the agent’s perspective.

---

Why Remote Tools?

• Python ML models that are impractical to run in-JVM
• Existing microservices you want to expose as agent capabilities
• Polyglot teams where tools are maintained in different languages
• Sandboxing computationally intensive or security-sensitive operations

---

ProcessAgentTool

Executes an external subprocess and communicates via the AgentEnsemble subprocess protocol.

Subprocess Protocol Specification

Communication is JSON-over-stdio:

Input — written to the process’s stdin immediately after launch:

{"input": "the agent's input string"}

Success output — the process writes to stdout before exiting with code 0:

{"output": "result text for the LLM", "success": true}

Optionally include a structured payload for programmatic consumers:

{"output": "result text for the LLM", "success": true, "structured": {"key": "value"}}

Failure output — on a logical failure (not an exception):

{"error": "description of what went wrong", "success": false}

Non-zero exit code — treated as a failure; stderr is captured as the error message.

Timeout — configurable; the process is killed with destroyForcibly() if exceeded.

Note: Processes that do not read stdin (e.g., echo) are supported. The stdin write failure is logged at DEBUG level and execution continues normally.

Usage

var sentiment = ProcessAgentTool.builder()
    .name("sentiment_analysis")
    .description("Analyzes the sentiment of a piece of text. Returns positive, negative, or neutral.")
    .command("python3", "/opt/tools/sentiment.py")
    .timeout(Duration.ofSeconds(30))
    .build();

var agent = Agent.builder()
    .role("Analyst")
    .goal("Analyze customer feedback")
    .tools(List.of(sentiment))
    .llm(chatModel)
    .build();

Example Python Implementation

import sys
import json

def analyze(text):
    # Replace with real sentiment logic
    words = text.lower().split()
    if any(w in words for w in ["good", "great", "excellent"]):
        return "positive"
    elif any(w in words for w in ["bad", "terrible", "awful"]):
        return "negative"
    return "neutral"

data = json.loads(sys.stdin.read())
result = analyze(data["input"])
print(json.dumps({"output": result, "success": True}))

Builder Options

Option	Required	Default	Description
name	Yes	-	Tool name shown to the LLM
description	Yes	-	Tool description shown to the LLM
command	Yes	-	Program and arguments (varargs or List)
timeout	No	30 seconds	Maximum execution time before process kill

---

HttpAgentTool

Wraps an HTTP endpoint as an agent tool. No subprocess overhead — uses Java’s built-in HttpClient.

Request Format

• GET requests: input appended as a query parameter — ?input=<url-encoded-input>
• POST requests: input sent as the request body
  • Content-Type: application/json when input is valid JSON
  • Content-Type: text/plain; charset=UTF-8 otherwise

Response Format

The HTTP response body is returned as the tool’s plain-text output to the agent. HTTP 4xx/5xx responses are treated as failures.

Usage

GET endpoint (simple search API)

var search = HttpAgentTool.get(
    "knowledge_base_search",
    "Searches the internal knowledge base for relevant articles.",
    "https://kb.example.com/search");

POST endpoint with custom headers

var classifier = HttpAgentTool.builder()
    .name("text_classifier")
    .description("Classifies text into categories: tech, finance, sports, other.")
    .url("https://ml.example.com/classify")
    .method("POST")
    .header("Authorization", "Bearer " + System.getenv("ML_API_KEY"))
    .timeout(Duration.ofSeconds(60))
    .build();

POST endpoint with JSON input

// When the agent provides JSON input, application/json Content-Type is set automatically
var enricher = HttpAgentTool.post(
    "entity_enricher",
    "Enriches entity data given a JSON object with 'name' and 'type' fields.",
    "https://api.example.com/enrich");

Builder Options

Option	Required	Default	Description
name	Yes	-	Tool name shown to the LLM
description	Yes	-	Tool description shown to the LLM
url	Yes	-	The HTTP endpoint URL
method	No	POST	HTTP method (GET, POST, PUT, etc.)
header	No	(none)	Add a request header (can be called multiple times)
timeout	No	30 seconds	Request timeout

Example Node.js Service

const express = require('express');
const app = express();
app.use(express.text());

app.post('/classify', (req, res) => {
    const input = req.body;
    const category = classify(input);  // your logic here
    res.send(category);
});

app.listen(8080);

---

Mixing Native and Remote Tools

Tools of all types can be combined freely on a single agent:

var agent = Agent.builder()
    .role("Data Analyst")
    .goal("Analyze sales data using all available tools")
    .tools(List.of(
        new CalculatorTool(),                                   // Java (in-process)
        ProcessAgentTool.builder()                             // Python subprocess
            .name("forecast")
            .description("Generates a sales forecast")
            .command("python3", "/opt/tools/forecast.py")
            .build(),
        HttpAgentTool.get(                                     // REST API
            "product_data",
            "Retrieves product catalog data",
            "https://api.example.com/products")
    ))
    .llm(chatModel)
    .build();

---

Parallel Execution

When the LLM requests multiple tools in a single turn, AgentEnsemble executes them concurrently using the configured tool executor (default: virtual threads). This is especially beneficial for remote tools where I/O latency dominates:

// A parallel tool turn (calculator + Python forecast + HTTP API) runs concurrently
Ensemble.builder()
    .agent(agent)
    .task(task)
    .toolExecutor(Executors.newVirtualThreadPerTaskExecutor()) // default
    .build()
    .run();

For rate-limited APIs, provide a bounded executor to cap concurrency:

Ensemble.builder()
    .agent(agent)
    .task(task)
    .toolExecutor(Executors.newFixedThreadPool(4))
    .build()
    .run();