Clickable product demo

Debug an AI agent run like an execution trace.

This demo is intentionally self-contained: no login, no backend, no live model calls. It shows the product story that matters in an interview: trace capture, runtime enforcement, deterministic replay, and mismatch detection.

Policy config

PolicyConfig(max_steps=20, max_tokens=50000, block_duplicate_input_hashes=True)

Scenario 1

Watch a normal agent run become an execution record.

Completed

A support triage agent classifies a ticket, queries internal data, checks policy thresholds, and writes a replayable trace.

Steps used

0 / 20

Tokens used

0 / 50,000

Observed latency

0 ms

run_7f31a92c

Rows stream in at 150ms each. Click a row to inspect the full event JSON.

0% streamed

Agent wrapper

runtime = PaprikaRuntime(
    policy=PolicyConfig(
        max_steps=20,
        max_tokens=50000,
        block_duplicate_input_hashes=True,
    )
)

@runtime.agent(name="support_triage")
def triage(ctx, ticket):
    plan = ctx.llm.call(provider="openai", model="gpt-4.1-mini", input=ticket)
    account = ctx.tools.call("sql_query", {"account_id": plan.account_id})
    return ctx.llm.call(provider="anthropic", model="claude-sonnet-4-5", input=account)

Selected step

#1 / run_start / 4d7a91c2

Paprika opens an execution record before the agent does any work. This gives the run a stable id and a place to attach every later LLM call, tool call, policy check, and final status.

{
  "step": 1,
  "timestamp": "10:42:01.004",
  "type": "run_start",
  "label": "support_triage started",
  "inputHash": "4d7a91c2",
  "tokens": 0,
  "latencyMs": 3,
  "status": "ok",
  "explanation": "Paprika opens an execution record before the agent does any work. This gives the run a stable id and a place to attach every later LLM call, tool call, policy check, and final status.",
  "data": {
    "agent_name": "support_triage",
    "run_id": "run_7f31a92c",
    "policy": "standard_interview_demo"
  }
}