Demo

pytest --agentreplay=record tests/test_refund.py

RECORD  tests/cassettes/test_refund/test_refund_flow.jsonl
  ✓ Intercepted via PydanticAI capability hooks
  ✓ 2 model request/response turns captured
  ✓ 3 tool calls recorded (lookup_customer, check_refund_policy, refund_customer)
  ✓ Canonical JSONL cassette saved (format_version=1)

tests/test_refund.py::test_refund_flow PASSED [100%]
====================== 1 passed in 2.45s (recorded to disk) ======================

pytest --agentreplay=replay tests/test_refund.py

REPLAY  tests/cassettes/test_refund/test_refund_flow.jsonl
  ✓ ZERO live model API calls (skipped via SkipModelRequest)
  ✓ ZERO tool/network execution (skipped via SkipToolExecution)
  ✓ Replay position cursor matched all 7 cassette events
  ✓ 100% deterministic test execution

tests/test_refund.py::test_refund_flow PASSED [100%]
====================== 1 passed in 0.04s (60x faster / zero tokens used) ======================

pytest --agentreplay=replay tests/test_refund.py

FAILED tests/test_refund.py::test_refund_flow - DivergenceError:

Agent trajectory changed

Expected:
  1. model_request
  2. tool_call: lookup_customer(id='123')
  3. tool_call: check_refund_policy(tier='gold')
  4. tool_call: refund_customer(amount=39)
  5. model_response → "Refund processed."

Actual:
  1. model_request
  2. tool_call: lookup_customer(id='123')
  3. tool_call: refund_customer(amount=39)

Divergence at step 3:
  - tool_call: check_refund_policy(tier='gold')
  + tool_call: refund_customer(amount=39)
====================== 1 failed in 0.05s ======================

Overview

Testing AI agents in continuous integration is often painful:

• Live LLM calls in CI are slow, expensive, and flaky.
• Traditional mocks are brittle and easily miss subtle agent drifts (such as skipping a verification tool or altering argument payloads).
• Raw snapshot tests generate massive, noisy JSON diffs cluttered with timestamps, request IDs, and non-deterministic tokens.

pytest-agentreplay brings deterministic VCR-style testing to AI agents:

1. Record once against live models and tools during local test development.
2. Replay offline in CI with zero network and zero model API calls.
3. Catch behavioural drift with step-by-step trajectory diffs whenever tools, arguments, or execution sequences change.

[!NOTE] pytest-agentreplay is a testing tool, not an agent runtime or orchestration system. You don't need to rewrite your agent or replace your framework runtime.

Features

• ⚡ Zero-API Replay: Substitutes model responses and tool executions offline — test suites execute in milliseconds.
• 🔍 Structural Trajectory Diffs: Highlights the exact step where an agent diverged instead of dumping raw JSON walls.
• 🎯 Non-Invasive Adapter: Integrates with PydanticAI via public capability hooks without monkey-patching HTTP clients.
• 🛡️ No Silent Fallback: Replay never secretly falls back to live network calls; cassette exhaustion and unexpected tool calls fail loudly.
• 📦 Git-Friendly Cassettes: Canonical JSONL serialization (sorted keys, compact separators) produces byte-identical files across operating systems.
• 🧪 Pytest-Native: Integrated --agentreplay CLI option and test fixture for seamless workflow switching.

Installation

Install pytest-agentreplay using uv or pip:

# Core package
uv add pytest-agentreplay

# With specific framework adapters
uv add pytest-agentreplay[pydantic-ai]  # PydanticAI
uv add pytest-agentreplay[openai]       # OpenAI Python SDK
uv add pytest-agentreplay[anthropic]    # Anthropic Python SDK

# Install all adapters
uv add pytest-agentreplay[all]

Quick Start

1. Using the pytest Fixture (Recommended)

Add the agentreplay fixture to your existing test function. The cassette path is automatically derived from the test module and function name:

from your_app import support_agent


def test_refund_flow(agentreplay):
    caps = [c for c in [agentreplay.capability()] if c is not None]
    result = support_agent.run_sync("Refund order 123", capabilities=caps)

    assert "refund" in result.output.lower()

Step 1: Record Real Interactions

Run pytest with --agentreplay=record to capture live model and tool events into a cassette:

pytest --agentreplay=record tests/test_refund.py

RECORD  tests/cassettes/test_refund/test_refund_flow.jsonl
✓ model interactions recorded
✓ tool interactions recorded

Step 2: Replay Offline in CI

Run with --agentreplay=replay to run tests entirely offline:

pytest --agentreplay=replay tests/test_refund.py

REPLAY  tests/cassettes/test_refund/test_refund_flow.jsonl
✓ zero model API calls
✓ zero network
✓ deterministic

[!TIP] Running pytest without the --agentreplay flag executes tests normally without recording or intercepting interactions.

2. Programmatic Usage

You can also control recording and replaying explicitly in code without pytest CLI flags:

import agentreplay
from your_app import support_agent


def test_custom_refund():
    result = support_agent.run_sync(
        "Refund order 123",
        capabilities=[
            agentreplay.pydantic_ai(
                mode="replay",  # or "record"
                cassette_path="tests/cassettes/custom_refund.jsonl",
            )
        ],
    )
    assert "refund" in result.output.lower()

from pathlib import Path
from pydantic_ai import Agent
from pydantic_ai.models.test import TestModel
from pydantic_ai.tools import RunContext
import agentreplay


# 1. Define your agent & tools
def lookup_customer(ctx: RunContext[None], customer_id: str) -> str:
    return f"Customer {customer_id}: Alice (tier=gold)"


support_agent = Agent(
    TestModel(call_tools=["lookup_customer"]),
    tools=[lookup_customer],
)

cassette = Path("test_refund.jsonl")

# 2. Record: live interaction captured to JSONL
cap_record = agentreplay.pydantic_ai(mode="record", cassette_path=cassette)
record_result = support_agent.run_sync("Find customer 123", capabilities=[cap_record])
print("Recorded output:", record_result.output)

# 3. Replay: 100% offline, zero model/tool execution
cap_replay = agentreplay.pydantic_ai(mode="replay", cassette_path=cassette)
replay_result = support_agent.run_sync("Find customer 123", capabilities=[cap_replay])
print("Replayed output:", replay_result.output)
assert record_result.output == replay_result.output

Behavioural Trajectory Diffs

When an agent changes its decision path (e.g. prompt changes, tool parameter updates, or model upgrade drifts), agentreplay provides a clear, numbered trajectory diff:

FAILED tests/test_refund.py::test_refund_flow - DivergenceError:

Agent trajectory changed

Expected:
  1. model_request
  2. tool_call: lookup_customer(id='123')
  3. tool_call: check_refund_policy(tier='gold')
  4. tool_call: refund_customer(amount=39)
  5. model_response → "Refund processed."

Actual:
  1. model_request
  2. tool_call: lookup_customer(id='123')
  3. tool_call: refund_customer(amount=39)

Divergence at step 3:
  - tool_call: check_refund_policy(tier='gold')
  + tool_call: refund_customer(amount=39)

Real-World Examples

Explore fully functional agent examples in the examples/ directory:

Supported Frameworks & Adapters

1. PydanticAI

import agentreplay

# Pass capability directly to Agent.run_sync or run
result = agent.run_sync(
    "User prompt",
    capabilities=[agentreplay.pydantic_ai(mode="replay", cassette_path="cassette.jsonl")],
)

2. OpenAI SDK (Sync & Async)

from openai import OpenAI
import agentreplay

# Wrap any OpenAI client
client = agentreplay.openai(OpenAI(), mode="replay", cassette_path="cassette.jsonl")
completion = client.chat.completions.create(model="gpt-4o", messages=[...])

3. Anthropic SDK (Sync & Async)

from anthropic import Anthropic
import agentreplay

# Wrap any Anthropic client
client = agentreplay.anthropic(Anthropic(), mode="replay", cassette_path="cassette.jsonl")
message = client.messages.create(model="claude-3-5-sonnet-20241022", messages=[...])

4. Custom Python Tools (@agentreplay.tool)

import agentreplay


@agentreplay.tool
def execute_payment(account_id: str, amount: float) -> dict:
    # On record: executes real code & saves result
    # On replay: skips execution & returns recorded result
    return {"status": "success", "tx_id": "tx_999"}

How It Works

┌────────────────────────────────────────────────────────┐
│                      Agent Test                        │
└───────────────────────────┬────────────────────────────┘
                            │
              ┌─────────────┴─────────────┐
              ▼                           ▼
      [Record Mode]               [Replay Mode]
              │                           │
     Live Model & Tools           Cassette JSONL File
              │                           │
  Intercepts via Capability       Intercepts via Hooks:
    • Model Requests/Responses      • SkipModelRequest
    • Tool Calls/Results            • SkipToolExecution
              │                           │
              ▼                           ▼
   Saves Canonical JSONL         Zero Network / API Calls

• Record Mode: Intercepts model requests, model responses, and tool executions via PydanticAI's AbstractCapability hooks. Deep-copies all arguments and results to prevent mutation side-effects.
• Replay Mode: Uses PydanticAI's SkipModelRequest to return recorded responses without model invocations, and SkipToolExecution to substitute recorded tool outputs.
• Divergence Engine: Tracks execution position with a stateful cursor. Rejects mismatches in event kinds, unexpected tool names, altered arguments, cassette exhaustion, and unconsumed leftover events.

Cassette Format

Cassettes are stored as streamable, Git-diffable JSON Lines (JSONL) files. Each interaction turn is appended in real-time as a canonical TraceEvent.

Line 1 contains the cassette header with format versioning and metadata:

{"created_at":"2026-08-27T08:30:00+00:00","format_version":1,"framework":"pydantic-ai"}

Subsequent lines contain canonical TraceEvent records:

{"event_id":"a1b2c3d4e5f6","kind":"run_start","timestamp":1724747400.0}
{"event_id":"b2c3d4e5f6a1","kind":"model_request","name":"gpt-4o","timestamp":1724747401.0}
{"arguments":{"customer_id":"123"},"event_id":"c3d4e5f6a1b2","kind":"tool_result","name":"lookup_customer","result":{"tier":"gold"},"timestamp":1724747402.0}
{"event_id":"d4e5f6a1b2c3","kind":"model_response","result":{"parts":[{"content":"Processed.","part_kind":"text"}]},"timestamp":1724747403.0}
{"event_id":"e5f6a1b2c3d4","kind":"run_end","timestamp":1724747404.0}

[!IMPORTANT] Serialization is deterministic: dictionary keys are sorted, compact separators are enforced, and encoding is UTF-8. Logically identical traces produce byte-identical files across platforms.

CLI & Fixture Reference

Pytest CLI Flags

agentreplay Fixture Methods

• agentreplay.mode — Returns "record", "replay", or None.
• agentreplay.capability(cassette_path=None) — Creates an AgentReplayCapability instance configured with the active mode and target cassette path.
• agentreplay.default_cassette_path — Returns the auto-derived path: tests/cassettes/{module_name}/{test_name}.jsonl.

Development

Set up a local development environment with uv:

# Clone the repository
git clone https://github.com/aafre/agentreplay.git
cd agentreplay

# Install dependencies
uv sync --all-extras

# Run quality gates
uv run ruff check .
uv run ruff format --check .
uv run mypy .
uv run pytest -v