Name: Agent Trajectory Failure Forensics
Brand: Tetrees AI
Availability: InStock

Agent Trajectory Failure Forensics cover

Agent Trajectory Failure Forensics preview 2

Agent Trajectory Failure Forensics preview 3

Tetrees Agent guide

What this agent does

Long agent runs fail through interacting plans, tools, policies, and observations; reviewing only the final answer confuses the first causal failure with later symptoms and produces vague remediation.

Designed for

Agent-platform engineers
Evaluation and reliability teams
MCP tool authors
Teams reviewing failed automated workflows

Inputs it works with

Agent trajectory with messages, tool calls, and results
Task contract and domain policy
Tool schemas and success criteria
Optional known outcome or reviewer notes

Outputs you receive

Canonical trajectory map
Static and step-level invariant ledger
Critical-failure localization
Failure taxonomy and causal chain
Remediation and regression-test specification

How the workflow runs

1Normalize heterogeneous logs into a stable ordered trajectory without changing speaker or tool semantics.
2Derive static task, policy, tool, and structural invariants.
3Evaluate context-sensitive invariants at every step and record evidence for violations.
4Separate candidate failures from consequences and localize the earliest critical step.
5Classify the failure with a grounded taxonomy and calibrated confidence.
6Produce a replayable report, targeted regression case, and remediation priority.

Example usage

Start with one of these requests, then replace the details with your own.

Find why a tool-using support agent failed

A customer-service run ends with the wrong order change after several apparently reasonable tool calls.

Example request

Analyze this trajectory against the attached policy and tool schemas. Identify the first critical failure, distinguish downstream errors, classify it, and propose one regression case.

Expected outcome: A normalized trace, invariant violations with step evidence, earliest causal failure, taxonomy label, downstream chain, and a targeted regression test.

Compare two agent versions on the same task

A new release reaches the right final answer but violates policy during execution.

Example request

Compare these two trajectories. Check plan-policy alignment, tool arguments, observation interpretation, recovery, and whether final-answer success hides an unsafe intermediate step.

Expected outcome: A step-aligned comparison, hidden policy violation, causal attribution, confidence, and release-blocking criteria.

A successful run should

✓ Trajectory order and speaker identity remain intact
✓ Every failure cites a step and invariant
✓ The first causal failure is separated from consequences
✓ Taxonomy labels match evidence
✓ Remediation produces a testable regression

Boundaries and limitations

• Missing or redacted tool results can block confident localization
• An LLM judge remains fallible and needs sampled human review
• Outcome success does not prove policy compliance
• The pack diagnoses traces; it does not rerun external systems

Capabilities

trajectory normalizationinvariant analysisfailure localizationtaxonomy classificationremediation design

Choose OpenAI, Claude or Z.AI in Agent Studio, then run with Tetrees Points or BYOK.

Execution tools

Hosted runs use only this signed set. Local MCP may add tools that you separately configure and approve.

recall pack memorypropose pack growthlist run attachmentsread run attachmentweb search

Agent AVCP

Agent AVCP report

Tested v1.0.0

The tested TAIP artifact is bound to this report and platform signature. Scores describe tested evidence, not a guarantee of every future model response.

Overall score

9.5

out of 10

Security9.5

Function10.0

Completion10.0

Mandatory gates

provenance✓ Pass

costCeilings✓ Pass

growthRollback✓ Pass

skillExecution✓ Pass

Reviews

No reviews yet.

Agent Trajectory Failure Forensics