ROOT-CAUSE ANALYSIS FOR AI AGENTS
When your AI agent fails,
know exactly why.
Root-cause analysis for semantic failures in production agents. Forensi reads the trace you already have, identifies what went wrong, and shows the evidence and the likely fix.
CaseCASE-184 · RUN-4821
WORKED EXAMPLE · REPLAYED LOCALLY
Left and right arrow keys step this case.
Customer
Can I get a refund for the subscription I bought six days ago?
Evidence
Not yet read
Retrieved context has not been read yet.
A green trace is not a correct outcome.
- The run completed.
- The tools returned.
- The customer still got the wrong policy.
Traditional observability tells you what ran. Forensi investigates why the outcome was semantically wrong: which evidence was stale, which source should have won, and which evidence actually drove the answer.
What Forensi detects
- D1
- Grounding
- Are the output claims supported by the evidence the agent actually saw?
- D2
- Tool semantic consistency
- Did the agent interpret the tool result correctly?
- D3
- Cross-source conflict
- Do retrieved sources make incompatible claims about the same thing?
- D4
- Freshness / authority
- Did stale or lower-authority evidence override the current source of truth?
- D5
- Semantic tool status
- Did a technically successful tool call return something semantically incomplete, irrelevant or misleading?
Start with the traces you already have.
Reads
- messages
- retrieved context
- tool inputs
- tool outputs
- final response
- optional outcome
Returns
- failure classification
- evidence
- causal chain
- likely remediation
Works from the traces your agent already produces — no new retrieval stack required.
Reliability audit
Have an agent failing in production?
Bring us 20–50 failed runs.
We’ll return a reliability report showing root causes, recurring failure patterns, and recommended fixes.
Book a reliability audit