See Why Your Agents Fail

Not just errors—the chain of decisions that led there. What the LLM saw, what it decided, and why it went wrong.

Start Free View on GitHub
Two lines to start tracing
from agent_tracer import Tracer

tracer = Tracer(api_key="at_live_...")
tracer.instrument_anthropic() # or instrument_openai()

with tracer:
    response = client.messages.create(...)
    # ↑ This call is now traced with full I/O
support-agent Session 3m ago
Failed
llm_call analyze_request 847ms
Input:
{"message": "I want a refund for order #1234"}
Output:
{"intent": "refund", "order_id": "1234", "action": "lookup_order"}
tool_call lookup_order 124ms
Input:
{"order_id": "1234"}
Output:
{"status": "delivered", "total": 89.99, "item": "Wireless Headphones"}
llm_call decide_action 623ms
tool_call process_refund error 52ms
RefundPolicyError
Refund denied: order delivered >30 days ago (47 days)
Input:
{"order_id": "1234", "amount": 89.99}
💡
The error alone is useless. What matters: the LLM had the order date (47 days ago) but still decided to process the refund. Your prompt didn't tell it to check the policy window.

Click any event to inspect inputs and outputs

Most Agent Failures Aren't Exceptions

Your agent "succeeds" but does the wrong thing. No error thrown, just a bad decision you can't explain.

79%
Of failures from prompts and specs—not code bugs
Source: MAST Research
~50%
Task completion rate—half of "working" agents silently fail
Source: arXiv
?
What inputs led to failure? What did the LLM "think"?
Without tracing, you'll never know

The Error Isn't the Answer. The Context Is.

Real scenarios where knowing the error wasn't enough. Click to see what actually mattered.

Code Review Agent misses SQL injection

The agent reviewed a PR and approved it, but missed a critical security flaw.

Support Agent enters infinite loop

Token usage spiked 10x. The agent kept retrying the same failed action.

Research Agent hallucinates sources

Generated report cited papers that don't exist. Looked authoritative but was fabricated.

Pricing

Free

$0

  • 10,000 events/month
  • 100 sessions/month
  • 7 day retention
Get Started

Pro

$49/mo

  • 1M events/month
  • 10,000 sessions/month
  • 30 day retention
Start Free Trial

Stop guessing. See the decisions.

Understand why your agents fail—even when there's no error. Free tier, no credit card.

Create Account