agentskit.js
Evals

Deterministic replay

Record once, replay forever. Test without hitting the network.

Record

import { createRecordingAdapter } from '@agentskit/eval'

const rec = createRecordingAdapter({
  inner: openai({ apiKey }),
  cassettePath: '.agentskit/cassettes/triage.jsonl',
})

Run your suite once with rec — every call captured.

Replay

import { createReplayAdapter } from '@agentskit/eval'

const replay = createReplayAdapter({
  cassettePath: '.agentskit/cassettes/triage.jsonl',
})

Use replay in CI — zero network, deterministic.

Time travel

import { createTimeTravelSession } from '@agentskit/eval'

const session = createTimeTravelSession({ cassettePath })
session.rewindTo(step)
session.override(step, { output: 'alternate response' })
const forked = session.fork()

Replay against different model

import { replayAgainst } from '@agentskit/eval'

const diff = await replayAgainst({
  cassettePath,
  adapter: anthropic(...),
})
✎ Edit this page on GitHub·Found a problem? Open an issue →·How to contribute →

On this page