Getting Started

Send your first trace to Observe in under 5 minutes. No SDK required — just an HTTP POST.

1

Create an account

Sign up at agentscrimmage.com/register. The Free tier includes 5,000 traces/month and 500 gateway requests/month — no credit card required.

2

Create an agent and get your API key

After logging in, create a new agent. Go to the Observe section and click Enable Observe. Your API key will be displayed on the Setup tab — it starts with as_live_.

Copy the key. You'll use it as a Bearer token in the Authorization header for all API calls.

export SCRIMMAGE_API_KEY="as_live_your_key_here"
3

Send your first trace

POST a trace to the batch endpoint. This example sends a single agent response trace:

cURL
curl -X POST https://agentscrimmage.com/api/observe/traces/batch \
  -H "Authorization: Bearer $SCRIMMAGE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "traces": [
      {
        "traceId": "t-001",
        "sessionId": "conv-1024",
        "type": "agent_response",
        "llm": {
          "provider": "anthropic",
          "model": "claude-sonnet-4",
          "outputMessage": "Your order has been shipped and should arrive in 3-5 business days.",
          "inputTokens": 45,
          "outputTokens": 18,
          "latencyMs": 820
        }
      }
    ]
  }'

A successful response returns 202 with the number of traces ingested:

Response
{
  "ingested": 1,
  "flagged": 0
}
4

Open your dashboard

Go to your agent's Observe dashboard. Your trace appears in the Conversations tab. The system automatically runs:

  • Tier 1 — instant regex checks for PII, injection, structural issues (<10ms)
  • Tier 2 — async AI evaluation for hallucination, compliance, quality
  • Tier 3 — full 30-dimension session scoring when a conversation completes

Issues appear on the Issues tab. Scores appear on the Analytics tab.

5

Send more trace types

There are 5 trace types. Use the right one for each event:

TypeWhen to use
llm_callRaw LLM completion (includes model, tokens, latency)
agent_responseFinal response sent to user (may include RAG context)
user_messageUser input (for conversation replay and context)
tool_callTool/function execution (name, args, result, success)
errorError event (provider timeout, rate limit, exception)

See the Trace Types reference for the full field list for each type.

6

Optional: Add user tracking

Include a userId field on any trace to enable per-user analytics on the Users tab:

{
  "traceId": "t-002",
  "sessionId": "conv-1024",
  "userId": "user-7291",
  "type": "agent_response",
  "llm": {
    "provider": "anthropic",
    "model": "claude-sonnet-4",
    "outputMessage": "Your order has been shipped."
  }
}

The userIdis opaque — use any string format (email, UUID, internal ID). It's never shared between agents or customers.