Nabla-X AgentSense

Know why your agents did what they did.

Extends Decision Intelligence's explainability engine from single models to multi-agent systems — tracing a decision back through every hand-off between agents and tools, for teams running agent workflows in production who need to know where, and why, something went wrong.

Illustrative trace — target output, not a real run

Router Agent→Classified request as "refund dispute", routed to Billing Agent

Billing Agent→Called refund_lookup tool, received partial record (missing order_date)

Billing Agent→Assumed order_date = today, proceeded with refund calculation⚠ flagged

Approval Agent→Approved refund based on Billing Agent's (incorrect) date assumption

AgentSense→Root cause: missing field silently defaulted instead of triggering a clarification step

Core Capabilities

Cross-Agent Trace Reconstruction

Follows a task across every agent and tool call it touched, not just the final agent that produced the output.

Hand-Off Point Analysis

Identifies exactly which hand-off between agents introduced an error, an assumption, or a dropped constraint.

Plain-Language Root Cause

Explains what went wrong in a sentence a non-engineer can act on, backed by the full technical trace underneath.

How AgentSense Would Work

Planned design — the trace being captured across a timeline. This pipeline does not exist yet.

Planned Capture Timeline

not yet built

Capture the Multi-Agent Trace

Every agent invocation, tool call, and hand-off is logged with its inputs, outputs, and reasoning.

RouterBilling×2Approval

Reconstruct the Decision Chain

Decision Intelligence's explainability engine, extended to walk the graph of hand-offs.

Surface the Root Cause

Pinpoints the hand-off where things went wrong, with a plain-language explanation attached.

Root-Cause Report (plain-language)

→

Team

How AgentSense Would Think

Zoomed into one hand-off from the trace above — the reasoning approach this concept is designed around, not a working inspector.

Billing Agent → Approval Agent hand-offflagged

illustrative — diffing logic not yet designed

Input State

order_id: "ord_8841"

amount: 42.00

order_date: — missing

diffed

Output State

order_id: "ord_8841"

amount: 42.00

order_date: "today" (assumed)

Verdict

A missing field was silently defaulted instead of triggering a clarification step — the kind of gap this concept is meant to surface automatically.

Graph, Not Log

Treats a run as a directed graph of hand-offs, so it can point to a specific edge, not just a timestamp.

Built on Decision Intelligence

Reuses Decision Intelligence's explainability engine rather than a new one, extended for multi-agent chains.

Actionable, Not Descriptive

A fix a team can make, not a forensic log dump — plain-language summary is the primary output.

Framework-Agnostic by Intent

Designed to observe whatever agent framework a team already runs. Scope not yet validated.

Concept Architecture

Planned Architecture

Nothing below is built. This describes the intended design and its dependency on Decision Intelligence's engine, which is itself not yet real.

Dependency: blocked on Decision Intelligence's engine, which doesn't exist yetillustrative trace structure — not real telemetry

start

hand-off sequence →

AgentSense Engine

Not yet designed

extends Decision Intelligence's explainability engine

Router Agent

Hand-off logging schema

Unspecified

Billing Agent · call 1

Framework integration points

Unspecified

Billing Agent · call 2

Root-cause scoring, reused from Decision Intelligence

Blocked on Decision Intelligence

Approval Agent

Multi-agent graph extension

Not started

Points at the fix: this is the hand-off the root-cause engine would name, not a vague “something went wrong.”

Hand-Off Tracing

Follows a task across every agent it touched, not just where it ended up.

Points at the Fix

Names the specific hand-off responsible, not a vague “something went wrong.”

Built on Decision Intelligence

No separate engine — extends the same explainability core once it exists.

Built on Decision Intelligence

Nabla-X AgentSense

Know why your agents did what they did.

Illustrative trace — target output, not a real run

Router Agent→Classified request as "refund dispute", routed to Billing Agent

Billing Agent→Called refund_lookup tool, received partial record (missing order_date)

Billing Agent→Assumed order_date = today, proceeded with refund calculation⚠ flagged

Approval Agent→Approved refund based on Billing Agent's (incorrect) date assumption

AgentSense→Root cause: missing field silently defaulted instead of triggering a clarification step

Core Capabilities

Cross-Agent Trace Reconstruction

Follows a task across every agent and tool call it touched, not just the final agent that produced the output.

Hand-Off Point Analysis

Identifies exactly which hand-off between agents introduced an error, an assumption, or a dropped constraint.

Plain-Language Root Cause

Explains what went wrong in a sentence a non-engineer can act on, backed by the full technical trace underneath.

How AgentSense Would Work

Planned design — the trace being captured across a timeline. This pipeline does not exist yet.

Planned Capture Timeline

not yet built

Capture the Multi-Agent Trace

Every agent invocation, tool call, and hand-off is logged with its inputs, outputs, and reasoning.

RouterBilling×2Approval

Reconstruct the Decision Chain

Decision Intelligence's explainability engine, extended to walk the graph of hand-offs.

Surface the Root Cause

Pinpoints the hand-off where things went wrong, with a plain-language explanation attached.

Root-Cause Report (plain-language)

→

Team

How AgentSense Would Think

Zoomed into one hand-off from the trace above — the reasoning approach this concept is designed around, not a working inspector.

Billing Agent → Approval Agent hand-offflagged

illustrative — diffing logic not yet designed

Input State

order_id: "ord_8841"

amount: 42.00

order_date: — missing

diffed

Output State

order_id: "ord_8841"

amount: 42.00

order_date: "today" (assumed)

Verdict

A missing field was silently defaulted instead of triggering a clarification step — the kind of gap this concept is meant to surface automatically.

Graph, Not Log

Treats a run as a directed graph of hand-offs, so it can point to a specific edge, not just a timestamp.

Built on Decision Intelligence

Reuses Decision Intelligence's explainability engine rather than a new one, extended for multi-agent chains.

Actionable, Not Descriptive

A fix a team can make, not a forensic log dump — plain-language summary is the primary output.

Framework-Agnostic by Intent

Designed to observe whatever agent framework a team already runs. Scope not yet validated.

Concept Architecture

Planned Architecture

Nothing below is built. This describes the intended design and its dependency on Decision Intelligence's engine, which is itself not yet real.

Dependency: blocked on Decision Intelligence's engine, which doesn't exist yetillustrative trace structure — not real telemetry

start

hand-off sequence →

AgentSense Engine

Not yet designed

extends Decision Intelligence's explainability engine

Router Agent

Hand-off logging schema

Unspecified

Billing Agent · call 1

Framework integration points

Unspecified

Billing Agent · call 2

Root-cause scoring, reused from Decision Intelligence

Blocked on Decision Intelligence

Approval Agent

Multi-agent graph extension

Not started

Points at the fix: this is the hand-off the root-cause engine would name, not a vague “something went wrong.”

Hand-Off Tracing

Follows a task across every agent it touched, not just where it ended up.

Points at the Fix

Names the specific hand-off responsible, not a vague “something went wrong.”

Built on Decision Intelligence

No separate engine — extends the same explainability core once it exists.