inngest-agent-evals

Inngest Agent Evals

Use this skill when the user wants to evaluate AI agents or AI workflows in production, add scoring, compare prompts/models/tools, group related runs, or debug why an agent outcome was good or bad.

Agent Evals is not a separate package. It is the production evaluation workflow built from Inngest functions, durable steps, scores, deferred scorers, sessions, traces, experiments, and Insights.

Canonical References

Check the public docs first when exact API details matter:

• Agent Evals overview: https://www.inngest.com/docs/learn/agent-evals
• Scoring guide: https://www.inngest.com/docs/features/inngest-functions/steps-workflows/scoring
• Deferred scoring guide: https://www.inngest.com/docs/features/inngest-functions/steps-workflows/deferred-scoring
• Step experiments guide: https://www.inngest.com/docs/features/inngest-functions/steps-workflows/step-experiments
• Sessions: https://www.inngest.com/docs/features/events-triggers/sessions
• Traces: https://www.inngest.com/docs/platform/monitor/traces
• Insights: https://www.inngest.com/docs/platform/monitor/insights
• Scoring reference: https://www.inngest.com/docs/reference/typescript/v4/functions/scoring
• group.experiment() reference: https://www.inngest.com/docs/reference/typescript/v4/functions/group-experiment
• Launch context: https://www.inngest.com/blog/introducing-agent-evals

Decision Flow

Start from the outcome, not the mechanism:

1. Identify the product or quality signal: helpful click, ticket resolution, conversion, guardrail pass, retrieval quality, model confidence, latency, cost, human review, or LLM-as-judge result.
2. Decide when the signal appears:
   • During the run: use direct scoring with step.score() or inngest.score().
   • After the run: use deferred scoring with createScorer() and defer(), or explicit inngest.score({ runId }).
3. Decide how humans will inspect related work:
   • Use meta.sessions for repeated high-cardinality identifiers such as conversation_id, ticket_id, agent_run_id, or import_id.
   • Use Insights for ad hoc historical analysis, low-cardinality filters, SQL aggregates, and broad investigation.
4. Decide whether to compare variants:
   • Use group.experiment() for prompt, model, provider, tool, workflow, or rollout comparisons.
   • Pass experimentRef to deferred or later scores when the score should be credited to the selected variant.
5. Preserve traceability:
   • Keep model calls, tool calls, waits, database writes, and scoring work in durable steps so traces explain why the score happened.

Setup Requirements

• Use TypeScript SDK v4 and install the latest SDK: npm install inngest@latest.
• Scoring and deferred scoring are beta APIs; verify current imports against the docs before shipping user-facing examples.
• step.score() requires scoreMiddleware() from inngest/experimental on the Inngest client.
• createScorer() is imported from inngest/experimental and the scorer must be registered in the serve handler with other functions.
• Sessions sent through the TypeScript SDK require v4.7.0 or later.
• Step experiments require v4.8.0 or later for group.experiment() and the experiment helper from inngest.

Client shape:

import { Inngest } from "inngest";
import { scoreMiddleware } from "inngest/experimental";

export const inngest = new Inngest({
  id: "support-agent",
  middleware: [scoreMiddleware()],
});

Sessions

Add sessions when events belong to a user flow, conversation, ticket, import, or agent task that someone will inspect repeatedly.

await inngest.send({
  name: "support/ticket.created",
  data: {
    ticketId: "tk_123",
    message: "I can't sign in.",
  },
  meta: {
    sessions: {
      ticket_id: "tk_123",
    },
  },
});

Rules:

• Use stable, non-secret, high-cardinality IDs.
• Keep the key generic, such as conversation_id; put the actual ID in the value.
• Pass sessions explicitly through step.invoke() and step.sendEvent() when downstream runs should join the same session.
• Do not use sessions for labels like environment: prod; use Insights for that style of filtering.

Direct Scoring

Use direct scoring when the score is known during the function run.

Good direct scores:

• guardrail pass/fail
• JSON validity
• retrieval confidence
• tool success
• model confidence
• inline LLM-as-judge result

export const answerTicket = inngest.createFunction(
  {
    id: "answer-ticket",
    triggers: [{ event: "support/ticket.created" }],
  },
  async ({ event, step }) => {
    const answer = await step.run("generate-answer", () =>
      generateAnswer(event.data.ticketId)
    );

    const passed = await step.run("check-answer", () => validateAnswer(answer));

    await step.score("score-answer-quality", {
      name: "answer-quality",
      value: passed,
    });

    return { answer, passed };
  }
);

Use stable score names. Changing a score name creates a separate metric. Score values must be finite numbers or booleans.

Deferred Scoring

Use deferred scoring when the useful signal arrives after the parent workflow finishes, such as user feedback, conversion, ticket reopen, retention, or a slow LLM-as-judge run.

import { createScorer } from "inngest/experimental";
import { z } from "zod";

export const feedbackScorer = createScorer(
  inngest,
  {
    id: "support-feedback-scorer",
    schema: z.object({ ticketId: z.string() }),
  },
  async ({ event, step }) => {
    const feedback = await step.waitForEvent("wait-for-feedback", {
      event: "support/feedback.received",
      timeout: "7d",
      if: `async.data.ticketId == '${event.data.ticketId}'`,
    });

    return {
      name: "user-feedback",
      value: feedback?.data.helpful ? 1 : 0,
    };
  }
);

Trigger it from the producing function:

async ({ event, step, defer }) => {
  const answer = await step.run("generate-answer", () =>
    generateAnswer(event.data.ticketId)
  );

  defer("score-feedback", {
    function: feedbackScorer,
    data: { ticketId: event.data.ticketId },
  });

  return { answer };
}

Guardrails:

• Register scorers with serve({ functions: [...] }).
• defer() is fire-and-forget; do not await it.
• The parent run is attributed automatically for deferred scorers.
• Return a default score or null when a signal times out, based on the product semantics.

Experiments

Use group.experiment() when comparing prompts, models, providers, tools, workflow rewrites, or operational settings against real traffic.

import { experiment } from "inngest";

const { result, variant, experimentRef } = await group.experiment(
  "answer-style",
  {
    variants: {
      concise: () => step.run("answer-concise", () => answerConcise(event.data)),
      detailed: () =>
        step.run("answer-detailed", () => answerDetailed(event.data)),
    },
    select: experiment.bucket(event.data.accountId, {
      weights: { concise: 50, detailed: 50 },
    }),
  }
);

Selection strategy:

• experiment.weighted() for run-level traffic splits.
• experiment.bucket(stableId, { weights }) when a user, account, or tenant should usually keep the same experience.
• experiment.custom() when assignment comes from a database, flag service, or rollout table.
• experiment.fixed() to force one variant during testing or after choosing a winner.

Rules:

• Each variant callback must call at least one step.* tool.
• Keep experiment IDs and variant names stable because they appear in traces.
• The selected variant is memoized for retries and replays.
• Persist experimentRef and the parent run ID if a later, separate process will score the selected variant.

When deferred scoring a variant, pass the ref:

defer("score-answer-feedback", {
  function: feedbackScorer,
  data: { ticketId: event.data.ticketId },
  experiment: experimentRef,
});

When scoring from a later run explicitly:

await inngest.score.experiment({
  name: "clickthrough",
  value: 1,
  experiment: experimentRef,
  runId: originalRunId,
});

Brownfield Migration

When adding Agent Evals to an existing app:

1. Search for existing agent/workflow outputs and where users act on them.
2. Find durable identifiers already in the domain: conversation ID, ticket ID, account ID, import ID, run ID, or recommendation ID.
3. Add meta.sessions at event producers before building dashboards around correlation.
4. Add one direct score where the outcome is already known.
5. Add a deferred scorer only after the signal event exists or can be emitted cleanly.
6. Add group.experiment() only around one isolated decision at a time.
7. Keep old prompt/model/tool behavior stable until scoring proves the new path.

Use inngest-brownfield-audit first when the repo has many candidate workflows. Use inngest-agents with this skill when the work is an AgentKit or durable agent workflow.

Anti-Patterns

• Starting with a prompt experiment before naming the outcome metric.
• Scoring with unstable names, strings, objects, NaN, or Infinity.
• Forgetting scoreMiddleware() and then assuming step.score() is missing.
• Running scorers in an external queue when the scorer needs durable waits.
• Keeping experiment assignment only in memory.
• Scoring an experiment variant from a later run without the original runId.
• Using sessions for low-cardinality labels or sensitive personal data.
• Leaving model/tool work outside steps, making traces unable to explain the score.

Verification

• Typecheck the Inngest client, functions, scorer schemas, and serve handler.
• Confirm scorers are registered with the serve endpoint.
• Test the producer emits events with expected meta.sessions.
• Test direct scores use stable names and finite number/boolean values.
• Test deferred scorer timeout behavior.
• For experiments, test that each variant callback contains step work and that later scoring receives experimentRef plus the original run ID when needed.
• If possible, run the Inngest dev server and inspect traces, sessions, scores, and experiment variant selection in the dashboard or through available CLI/API tooling.