trigger-agents

Trigger.dev を使用した AI エージェントのパターン

Trigger.dev の耐久性の高い実行を活用して、本番環境に対応した AI エージェントを構築します。

パターンの選択

Need to...                              → Use
─────────────────────────────────────────────────────
Process items in parallel               → Parallelization
Route to different models/handlers      → Routing
Chain steps with validation gates       → Prompt Chaining
Coordinate multiple specialized tasks   → Orchestrator-Workers
Self-improve until quality threshold    → Evaluator-Optimizer
Pause for human approval                → Human-in-the-Loop (waitpoints.md)
Stream progress to frontend             → Realtime Streams (streaming.md)
Let LLM call your tasks as tools        → ai.tool (ai-tool.md)

---

基本パターン

1. プロンプトチェーン（ゲート付きの逐次実行）

LLM 呼び出しを連鎖させ、ステップ間で検証します。中間出力に問題があれば、早い段階で失敗させます。

import { task } from "@trigger.dev/sdk";
import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";

export const translateCopy = task({
  id: "translate-copy",
  run: async ({ text, targetLanguage, maxWords }) => {
    // Step 1: Generate
    const draft = await generateText({
      model: openai("gpt-4o"),
      prompt: `Write marketing copy about: ${text}`,
    });

    // Gate: Validate before continuing
    const wordCount = draft.text.split(/\s+/).length;
    if (wordCount > maxWords) {
      throw new Error(`Draft too long: ${wordCount} > ${maxWords}`);
    }

    // Step 2: Translate (only if gate passed)
    const translated = await generateText({
      model: openai("gpt-4o"),
      prompt: `Translate to ${targetLanguage}: ${draft.text}`,
    });

    return { draft: draft.text, translated: translated.text };
  },
});

---

2. ルーティング（分類 → 振り分け）

低コストのモデルで分類してから、適切なハンドラーに振り分けます。

import { task } from "@trigger.dev/sdk";
import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";
import { z } from "zod";

const routingSchema = z.object({
  model: z.enum(["gpt-4o", "o1-mini"]),
  reason: z.string(),
});

export const routeQuestion = task({
  id: "route-question",
  run: async ({ question }) => {
    // Cheap classification call
    const routing = await generateText({
      model: openai("gpt-4o-mini"),
      messages: [
        {
          role: "system",
          content: `Classify question complexity. Return JSON: {"model": "gpt-4o" | "o1-mini", "reason": "..."}
          - gpt-4o: simple factual questions
          - o1-mini: complex reasoning, math, code`,
        },
        { role: "user", content: question },
      ],
    });

    const { model } = routingSchema.parse(JSON.parse(routing.text));

    // Route to selected model
    const answer = await generateText({
      model: openai(model),
      prompt: question,
    });

    return { answer: answer.text, routedTo: model };
  },
});

---

3. 並列化

batch.triggerByTaskAndWait を使用して、独立した LLM 呼び出しを同時に実行します。

import { batch, task } from "@trigger.dev/sdk";

export const analyzeContent = task({
  id: "analyze-content",
  run: async ({ text }) => {
    // All three run in parallel
    const { runs: [sentiment, summary, moderation] } = await batch.triggerByTaskAndWait([
      { task: analyzeSentiment, payload: { text } },
      { task: summarizeText, payload: { text } },
      { task: moderateContent, payload: { text } },
    ]);

    // Check moderation first
    if (moderation.ok && moderation.output.flagged) {
      return { error: "Content flagged", reason: moderation.output.reason };
    }

    return {
      sentiment: sentiment.ok ? sentiment.output : null,
      summary: summary.ok ? summary.output : null,
    };
  },
});

参照： 高度なパターンについては references/orchestration.md をご覧ください。

---

4. オーケストレーターとワーカー（ファンアウト／ファンイン）

オーケストレーターが作業項目を抽出し、ワーカーに処理を分散して、結果を集約します。

import { batch, task } from "@trigger.dev/sdk";

export const factChecker = task({
  id: "fact-checker",
  run: async ({ article }) => {
    // Step 1: Extract claims (sequential - need output first)
    const { runs: [extractResult] } = await batch.triggerByTaskAndWait([
      { task: extractClaims, payload: { article } },
    ]);

    if (!extractResult.ok) throw new Error("Failed to extract claims");
    const claims = extractResult.output;

    // Step 2: Fan-out - verify all claims in parallel
    const { runs } = await batch.triggerByTaskAndWait(
      claims.map(claim => ({ task: verifyClaim, payload: claim }))
    );

    // Step 3: Fan-in - aggregate results
    const verified = runs
      .filter((r): r is typeof r & { ok: true } => r.ok)
      .map(r => r.output);

    return { claims, verifications: verified };
  },
});

---

5. 評価者とオプティマイザー（自己改善ループ）

生成 → 評価 → フィードバックを基に、承認されるまで再試行します。

import { task } from "@trigger.dev/sdk";

export const refineTranslation = task({
  id: "refine-translation",
  run: async ({ text, targetLanguage, feedback, attempt = 0 }) => {
    // Bail condition
    if (attempt >= 5) {
      return { text, status: "MAX_ATTEMPTS", attempts: attempt };
    }

    // Generate (with feedback if retrying)
    const prompt = feedback
      ? `Improve this translation based on feedback:\n${feedback}\n\nOriginal: ${text}`
      : `Translate to ${targetLanguage}: ${text}`;

    const translation = await generateText({
      model: openai("gpt-4o"),
      prompt,
    });

    // Evaluate
    const evaluation = await generateText({
      model: openai("gpt-4o"),
      prompt: `Evaluate translation quality. Reply APPROVED or provide specific feedback:\n${translation.text}`,
    });

    if (evaluation.text.includes("APPROVED")) {
      return { text: translation.text, status: "APPROVED", attempts: attempt + 1 };
    }

    // Recursive self-call with feedback
    return refineTranslation.triggerAndWait({
      text,
      targetLanguage,
      feedback: evaluation.text,
      attempt: attempt + 1,
    }).unwrap();
  },
});

---

Trigger 固有の機能

機能	可能になること	参照
Waitpoints	人による承認ゲート、外部コールバック	references/waitpoints.md
Streams	フロントエンドへのリアルタイムな進捗通知	references/streaming.md
ai.tool	LLM がタスクをツールとして呼び出せるようにする	references/ai-tool.md
batch.triggerByTaskAndWait	型安全な並列実行	references/orchestration.md

---

エラー処理

const { runs } = await batch.triggerByTaskAndWait([...]);

// Check individual results
for (const run of runs) {
  if (run.ok) {
    console.log(run.output);  // Typed output
  } else {
    console.error(run.error);  // Error details
    console.log(run.taskIdentifier);  // Which task failed
  }
}

// Or filter by task type
const verifications = runs
  .filter((r): r is typeof r & { ok: true } =>
    r.ok && r.taskIdentifier === "verify-claim"
  )
  .map(r => r.output);

---

クイックリファレンス

// Trigger and wait for result
const result = await myTask.triggerAndWait(payload);
if (result.ok) console.log(result.output);

// Batch trigger same task
const results = await myTask.batchTriggerAndWait([
  { payload: item1 },
  { payload: item2 },
]);

// Batch trigger different tasks (typed)
const { runs } = await batch.triggerByTaskAndWait([
  { task: taskA, payload: { foo: 1 } },
  { task: taskB, payload: { bar: "x" } },
]);

// Self-recursion with unwrap
return myTask.triggerAndWait(newPayload).unwrap();