AI agent debugger
Your agent failed. Find out why
in 30 seconds.
Paste a trace. Get the exact step that failed, what broke next, and how to fix it. Debug in the browser, record in code, and gate in CI.
trace_id
trc_8f2d94a1c3e7
agent
revenue-summarizer
total
6.93s
span_id
span_d4e5f6
tool
fetch_document
duration
4.82s
status
FAULT
input
{
"document_id": "rev_q3_2024",
"format": "markdown"
}error
TimeoutError: fetch_document exceeded 5000ms threshold
TOOL_TIMEOUT
Debugger output
Root Cause
fetch_document timed out after 4.8s. No retry policy is configured on this tool call. The agent proceeded without a valid document source.
What broke next
vector_search
returned 0 results — fallback activated
llm_completion
completed with stale cache — output degraded
Output
marked incomplete, missing source data
Suggested Fix
fetch_document:
timeout: 2000
retry:
attempts: 3
backoff: "exponential"
fallback_source: "cache_store"
No account · Paste any format · Instant results
What Sepurux catches
AI agents fail in specific ways. Most are invisible without a trace.
Sepurux scans your trace for these patterns. Each one maps to a root cause, not just an error code.
TOOL_TIMEOUT
Tool timed out, cascade started.
fetch_document hung at 4.8 s. The agent continued on empty output. Three affected steps failed silently.
SCHEMA_BREAK
Tool returned the wrong shape.
Expected string, got null. The LLM filled in the gap with plausible-sounding fabricated data.
RETRY_EXHAUSTED
Retried with identical params. All failed.
query_db was called three times with the same arguments. No backoff, no parameter change. Same error each time.
CONTEXT_OVERFLOW
Prompt hit the token limit silently.
16,420 tokens in, limit is 16,384. The model truncated the last 2,000 tokens without raising an error.
FALLBACK_TRIGGERED
Primary failed. Fallback returned nothing.
vector_search fell back to cache. Cache returned 0 results. The agent treated that as valid context.
HALLUCINATION_RISK
LLM completed from an empty tool output.
The document tool returned null. The model generated a summary anyway. Output looks real — it isn't.
How it works
Three steps. Usually under a minute.
01
Paste the trace
Drop in JSON, raw logs, a LangSmith run URL, or a bare trace ID. Sepurux parses whatever you have.
No SDK. No account. No setup.
02
Get the root cause
Not just the error — the step that caused it, what broke next, and which affected steps silently continued on bad data.
Scored by confidence and impact.
03
Apply the fix
Each failure surfaces a concrete code snippet — retry config, timeout tuning, schema guard, or context trim.
Replay it, stress-test it, or add it to CI when you're ready.
When you need to go deeper
Debug first. Go further when you need to.
Sepurux starts as a debugger. Then you can install the recorder, capture traces automatically, replay and stress-test the same path, and gate releases in CI.
Stress-test a passing trace
Take a trace that's working and inject failures — timeouts, schema breaks, rate limits. Find out if your agent handles them before your users do.
Run checks on a schedule
Run traces on a cron schedule and get notified when a passing agent starts failing. Catch regressions automatically.
Set limits on what your agent can do
Control which tools your agent can call. Block risky invocations or require human approval before sensitive tool calls go through.
Works with the stacks you already ship
Debug in the browser. Record in code. Gate in CI.
You have a broken agent.
Start here.
No account. No setup. Paste a trace and know what failed in 30 seconds.