TOOL CALLrefund_order
failedMissing required refund_id
Hosted private beta · invite access
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 →GitHub correlation
Compare failing and baseline revisions, rank relevant files and commits, then open an administrator-approved draft PR from an existing branch.
Connect GitHub →Verified recovery
Turn the failed run into a deterministic regression fixture and test a candidate outcome before release.
Verify an outcome →CLI and OTLP ingestion
Upload, diagnose, correlate, retrieve reports, or send protected OTLP/HTTP JSON traces from your delivery workflow.
View setup →Datadog import
Pull a bounded, read-only Datadog log window into a new incident while connector credentials stay server-side.
Import evidence →Optional AI refinement
Choose Gemini or OpenAI per incident. Model findings are accepted only when they cite supplied evidence IDs.
Refine a diagnosis →A 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.
Open the private beta 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.
Hosted private beta
Bring one failed run.
Leave with the first useful answer.
The hosted Doctor includes manual and Datadog evidence intake, CLI and OTLP ingestion, cited diagnosis, GitHub change correlation, optional AI refinement, regression fixtures, deterministic verification, audit activity, and administrator-controlled draft PRs.
Open AgentFixIt Doctor →Beta access code required