TOOL CALLrefund_order
failedMissing required refund_id
Private beta is open
Your AI agent broke.
Find out why.
AgentFixIt turns traces and logs into a cited root-cause diagnosis, a regression fixture, and a fix your team can verify before it ships.
Start with JSON, JSONL, OTLP, text logs, or Datadog. No agent SDK required.
Diagnosis ready92% confidence
The agent claimed success after the refund tool failed validation.
FIRST BAD DECISIONError state was ignored
root causeThe run continued without recovery.
FINAL RESPONSE“Your refund was issued.”
wrongContradicts tool result.
3 findings6 evidence citations1 regression fixture
Works with the stack you already have
OpenTelemetryDatadogGitHubJSON / JSONLText logs
Not another trace dashboard
Go from “what happened?”
to “here’s what to fix.”
Observability shows the run. AgentFixIt builds the causal case—without presenting an unsupported model guess as fact.
Grounded diagnosis
Find the earliest event that caused the outcome. Every factual claim links back to its evidence.
See a diagnosis →Code correlation
Connect the incident to relevant files and changes in GitHub, using read-only access by default.
Integrated betaVerified recovery
Turn the failed run into a deterministic regression fixture and test a candidate outcome before release.
Integrated betaA closed debugging loop
Evidence in. Confidence out.
Three focused steps take your team from a production complaint to a testable fix.
1
BRING THE RUN
Upload or ingest evidence
Start manually, send OTLP traces, or import a bounded Datadog log window. Secrets are redacted before analysis.
trace.production.jsonl24 events normalized
2
BUILD THE CASE
Locate the first bad decision
Deterministic rules identify contradictions and errors. Optional AI adds hypotheses only when it can cite the run.
evidence: evt_009
Tool failure was followed by an unsupported success claim.
3
PROVE THE FIX
Reproduce and verify
Create a regression fixture from the incident and validate that the proposed behavior handles the captured failure.
✓
Verification passed2/2 expectations satisfied
Designed for sensitive evidence
Your production data deserves a careful debugger.
Agent traces can contain customer messages, tool arguments, and credentials. The platform starts from least privilege, clear evidence boundaries, and human approval for write actions.
Try the local-first Doctor →Credentials stay server-side
Model, GitHub, and telemetry keys are never exposed to the browser.
Redaction before analysis
Common secrets are removed during normalization, before optional model refinement.
Citations over confidence theater
Facts, inferences, gaps, and source events remain visibly distinct.
Read-only by default
Repository writes require an explicit, separate action and stay auditable.
Developer preview
Bring one failed run.
Leave with the first useful answer.
The Doctor workflow is available now for local evaluation. Team workspaces, hosted ingestion, and automated pull-request workflows will roll out in stages.
Open AgentFixIt Doctor →No production deployment required