The lenses
Agent Chain lens
The Agent Chain lens reads a JSON trace from an agent system and locates the step in the chain where the structural shape changes. The output reads as a per-step structural attribution, so the buyer can pin the verdict to one step rather than the whole chain.
What it answers
The lens answers which step in an agent trace is responsible for the structural verdict on the final output. The answer is a per-step breakdown showing the structural reading at the input to each step and at the output of each step, with the step where the structural shape changes highlighted.
When to run it
Run the Agent Chain lens on any multi-agent or multi-step LLM pipeline where the final output has been diagnosed and the buyer needs step-level attribution. The lens is also used in incident review, where the buyer's team is investigating a misbehaving agent system and needs to know which step in the chain altered the structural shape.
Public-shaped output
The result page renders an Agent Chain panel with one row per step. Each row carries the step label, the structural reading at the input, the structural reading at the output, and a flag indicating whether this step is where the shape changed. A footer row reports the responsible step, namely the step whose output verdict drives the final verdict.
Input requirements
The lens consumes a JSON trace shaped like agent_trace_sample.json. The trace lists the steps in order, with each step carrying its input, its output, and any intermediate state the agent recorded. The lens engine parser is tolerant to common trace formats and surfaces a structural reading at each boundary.
Tier gate
Business tier and above unlock the lens. Free, Growth, and Pro do not.
| Tier | Unlocked | Steps supported |
|---|---|---|
| Free | ||
| Growth | ||
| Pro | ||
| Business | ✓ | Up to twenty |
| Enterprise | ✓ | Unbounded |
What buyers do with it
Buyers pin the structural break to one step rather than the whole chain. A multi-step agent system whose final output has a mixed or model-inadequate verdict often has one step that is doing the damage. The lens names that step, and the buyer's team rewrites or retrains it.
Reviewers reading the report under EU AI Act Article 9 (high-risk per Annex III, applicable from August 2026) treat the Agent Chain panel as documented per-step attribution; the panel reports the structural reading at each boundary and the responsible step, without prescribing a redesign.