// playground

See Retrace in 2 minutes

Replay a real AI-agent run step by step, watch Retrace catch the failure, then fork a step to prove the fix, the full record → replay → detect → fork loop, on curated sample tapes.

Synthetic sample runs, explore, then record your own.

Demo

Support agent invents a refund policy

A RAG billing agent retrieves the correct 30-day refund policy, then confidently tells the customer it's 90 days. A textbook grounding failure — caught and fixed.

completed4 steps994 tokens$0.0032.1s

Retrace detected 2 issues in this run

agent.planllm_callgpt-4o

Duration

690ms

Input tokens

320

Output tokens

Cost

$0.0011

Input

{
  "system": "You are a billing support agent. Answer the customer using the knowledge base.",
  "user": "Can I still get a refund? I bought the Pro plan 58 days ago."
}

Output

{
  "plan": [
    "search refund policy in KB",
    "answer with the policy window",
    "be concise"
  ]
}

Step 1 / 4

Do this on your own agents

One decorator records every LLM call, tool call, and error. Then you get this exact replay, MAST detection, and fork-to-prove-the-fix loop on your real traces.

import retrace

retrace.configure(api_key="rt_...")

@retrace.record(name="my-agent", resumable=True)
def run_agent(prompt: str):
    # your LLM + tool calls here, captured automatically
    return answer

import retrace retrace.configure(api_key="rt_...") @retrace.record(name="my-agent", resumable=True) def run_agent(prompt: str): # your LLM + tool calls here, captured automatically return answer