Going to Production

Best practices for deploying agents to production: error handling, observability, cost control, and security.

Moving from development to production means handling the things that don’t matter during prototyping: errors, costs, observability, and security. This guide covers the patterns that production agent deployments need.

Error Handling

Agent-Level Errors

The direct AgentRuntime result reports status; the host-facing AgentResult used by core reports finishReason.

const result = await agent.run("Do something complex")

if (result.status === "failed") {
  // Log the error and alert
  logger.error("Agent run failed", {
    iterations: result.iterations,
    error: result.error,
  })
}

Tool Errors

Tools should return errors as data rather than throwing. This lets the agent see the error and adjust:

const apiTool: AgentTool<{ payload: unknown }, { success: boolean; data?: unknown; error?: string }> = {
  name: "call_service",
  // ...
  execute: async (input) => {
    try {
      const result = await service.call(input)
      return { output: { success: true, data: result } }
    } catch (err) {
      return {
        output: { success: false, error: err.message },
        isError: true,
      }
    }
  },
}

Mistake Limits

Core sessions can track consecutive recoverable mistakes and stop with finishReason: "mistake_limit". Direct runtime runs report failure through status and error.

Loop Detection

Loop detection is available through core/session execution settings.

Observability

OpenTelemetry Integration

The SDK integrates with OpenTelemetry for traces, metrics, and logs:

const cline = await ClineCore.create({
  clientName: "production-app",
  telemetry: myTelemetryService,
  logger: myLogger,
})

Events emitted include:

• agent_created — agent initialized with config
• tool_usage — tool called with name, duration, success/failure
• model_api_call — LLM API call with model, tokens, latency
• session_ended — session completed with finish reason, total tokens

Custom Metrics via Plugins

const productionMetrics: AgentPlugin = {
  name: "production-metrics",
  manifest: { capabilities: ["hooks"] },

  onRunEnd({ result }) {
    metrics.histogram("agent.duration_ms", result.durationMs)
    metrics.counter("agent.tokens.input", result.usage.inputTokens)
    metrics.counter("agent.tokens.output", result.usage.outputTokens)
    metrics.counter(`agent.finish.${result.finishReason}`, 1)
  },

  onToolCall({ call }) {
    metrics.counter(`agent.tool.${call.name}`, 1)
  },

  onError({ error, recoverable }) {
    metrics.counter("agent.errors", 1, {
      recoverable: String(recoverable),
      type: error.constructor.name,
    })
  },
}

Structured Logging

import pino from "pino"

const logger = pino({ level: "info" })

const agent = new Agent({
  // ...config
})

agent.subscribe((event) => {
  if (event.type === "run-finished") {
    logger.info({
      event: "agent_complete",
      status: event.result.status,
      iterations: event.result.iterations,
      tokens: event.result.usage,
    })
  }
  if (event.type === "run-failed") {
    logger.error({
      event: "agent_error",
      message: event.error.message,
    })
  }
})

Cost Control

Set Token Limits

const agent = new Agent({
  // ...config
  maxTokensPerTurn: 4096,
  maxIterations: 20,
})

Track Spending

type ModelPricing = {
  inputPerMillion: number
  outputPerMillion: number
}

function estimateCost(
  usage: { inputTokens: number; outputTokens: number; totalCost?: number },
  pricing: ModelPricing,
) {
  if (usage.totalCost != null) {
    return usage.totalCost
  }

  return (
    (usage.inputTokens / 1_000_000) * pricing.inputPerMillion +
    (usage.outputTokens / 1_000_000) * pricing.outputPerMillion
  )
}

let sessionCost = 0

const providerId = "anthropic"
const modelId = "claude-sonnet-4-6"
// Load current USD-per-million-token pricing from configuration or a pricing service.
const pricing = loadModelPricing(providerId, modelId)

const agent = new Agent({
  providerId,
  modelId,
  // ...
})

agent.subscribe((event) => {
  if (event.type === "usage-updated") {
    sessionCost = estimateCost(event.usage, pricing)

    if (sessionCost > 1.0) {
      agent.abort("Cost limit exceeded ($1.00)")
    }
  }
})

Use Cheaper Models for Simple Tasks

// Quick lookup: use a fast, cheap model
const lookupAgent = new Agent({
  providerId: "anthropic",
  modelId: "claude-haiku-4-5",
  // ...
})

// Complex work: use a more capable model
const workAgent = new Agent({
  providerId: "anthropic",
  modelId: "claude-sonnet-4-6",
  // ...
})

Security

Sandbox Tool Execution

Never let an agent run arbitrary shell commands on a production server without constraints:

const sandboxedBash: AgentTool<{ command: string }, { error?: string; result?: unknown }> = {
  name: "run_commands",
  description: "Execute a shell command in a sandboxed environment.",
  inputSchema: {
    type: "object",
    properties: {
      command: { type: "string", description: "Shell command to execute" },
    },
    required: ["command"],
  },
  execute: async (input) => {
    const blocked = ["rm -rf", "sudo", "curl | sh", "wget | sh", "> /dev/"]
    if (blocked.some((b) => input.command.includes(b))) {
      return { output: { error: "Command blocked by security policy" }, isError: true }
    }

    // Execute in a container, chroot, or restricted shell
    return { output: { result: await executeInSandbox(input.command) } }
  },
}

Validate Tool Inputs

Don’t trust tool inputs any more than you’d trust user input:

const fileTool = createTool({
  name: "read_files",
  // ...
  execute: async (input) => {
    const resolved = path.resolve(input.path)

    // Prevent path traversal
    if (!resolved.startsWith(ALLOWED_DIRECTORY)) {
      return { error: "Access denied: path outside allowed directory" }
    }

    return await fs.readFile(resolved, "utf-8")
  },
})

Rotate API Keys

Use environment variables for API keys, not hardcoded values:

// Good
const agent = new Agent({
  apiKey: process.env.ANTHROPIC_API_KEY,
  // ...
})

// Never do this
const agent = new Agent({
  apiKey: "sk-ant-api03-...",
  // ...
})

Deployment Patterns

Stateless Worker

For request/response workflows (API endpoints, webhook handlers):

async function handleRequest(prompt: string): Promise<string> {
  const agent = new Agent({
    providerId: "anthropic",
    modelId: "claude-sonnet-4-6",
    apiKey: process.env.ANTHROPIC_API_KEY,
    systemPrompt: "You are a helpful assistant.",
    tools: [myTool],
    maxIterations: 10,
  })

  const result = await agent.run(prompt)
  // Agent has no explicit shutdown method
  return result.outputText
}

Persistent Service

For long-running services with session management:

const cline = await ClineCore.create({
  clientName: "my-service",
  backendMode: "hub",
})

// Sessions persist across restarts
// Connectors, schedules, and team state all managed by the hub

process.on("SIGTERM", async () => {
  await cline.dispose("Shutting down")
  process.exit(0)
})