Platform · Agent tracing
Understand the decisions behind every governed turn.
IntSen rebuilds what your agent did—conversations, turns, model calls, the tools the model asked for and the results it got back—from the traffic that passes through the gateway. No tracing SDK required.
Agent turn
Fleet Service Advisor · turn 2
- Duration
- 3.6s
- Calls
- 2
- Cost
- $0.0042
- Quality
- 92
Model asks for a tool
The model's response contains a tool call. IntSen reads it from the response body; the agent will run it.
- tool
- check_inventory
- call_id
- call_7f3a
- arguments
- { part: "PS-4410 pump seal", site: "Dallas" }
- finish_reason
- tool_calls
- output tokens
- 46
Illustrative product scenario — synthetic data
Use ↑ ↓ to move between steps
The problem
A list of API calls doesn't tell you what the agent did.
Agents call the model many times per user message. Without grouping, you cannot see a turn as one unit of work, what it cost, which tools it used or whether it finished.
- Calls are not tied to the user message that caused them
- Tool use and loops are invisible in provider logs
- Cost and safety results can't be read per turn
Capabilities
What Agent tracing does
Conversations and turns
Tool calls and results
Turn completion
Sub-agents
Cost and checks per turn
Streaming aware
Technical detail
Why the gateway is enough for the core view
In an agent loop the model is the decision maker. It answers with a tool call or a final answer; the agent runs the tool and sends the result back to the model. Each decision and most results therefore pass through the gateway.
- Model call
- each gateway request and response: model, tokens, cost, latency, finish reason
- Tool call
- tool name, arguments and call id from the model's response
- Tool result
- the next request's tool message, matched by call id
- Tool time
- approximate: from the response with the call to the request with its result
- Final answer
- a response with no tool call, confirmed by the grace window
- Missing result
- a tool call with no result by turn end is marked as such
Why it matters
For the people accountable for it
Engineering
Debug a turn step by step without adding a tracing library to every agent.
CISO / risk
Investigate what a governed agent asked for and received, with the policy verdicts alongside.
FinOps
See which turns, tools and loops drive cost.
Boundaries
What to know before you evaluate
We would rather you hear the limits from us than discover them in a pilot.
- IntSen does not see steps the model did not request—for example retrieval your agent runs before the first model call. Retrieved content is visible only as part of the prompt.
- Tool executions themselves are not observed, and actions whose results are never returned to the model (an email sent, a ticket created) are not visible.
- Retries handled in your code, the exact answer shown to the user if your app post-processes it, and calls that bypass the gateway are not visible.
- Per-tool timing for parallel tool calls is approximate. Optional OpenTelemetry enrichment to close these gaps is on the roadmap.
FAQ
Common questions
What if we forget the task id?
Calls still work and are recorded, but each call becomes its own turn. The Console flags agents that send calls without a task id.
Do we tell IntSen when a turn ends?
No. IntSen sees the model's final answer and closes the turn itself after a short grace window.
What if body capture is off?
You still get structural steps—tool names, call ids and timings—but no arguments or results.
Ready to make your enterprise AI accountable?
See how IntSen can help your teams govern access, safety, quality and spending across the AI applications and agents you already run.