AI TrailMaps
AI TrailMaps draws a map of what an agent actually did, built from its own traces: every model it called, every tool and MCP server it used, the sub-agents and steps inside it, and the datastores and external services it reached. Use it to answer "what does this agent depend on?", "where does data leave the organization?" and "where is it failing?".
Enterprise feature. Open it from Discover AI → AI TrailMaps. It needs the AI TrailMaps permission, which Owner, Department Admin, Security & Compliance and Auditor have. Custom roles get it under the Discover AI heading in the role builder. AI TrailMaps must also be enabled for your deployment; if it is not in the sidebar, ask your ANTS Platform administrator. See Members & Roles.
How the map is built
A background job reads each agent's traces and stores one diagram per agent per day. The page combines the days in the window, which is always the last 7 days.
- Diagrams are rebuilt periodically, so new activity can take several hours to appear, depending on your deployment's schedule.
- The Data-Flow replay, trace list, sessions and failure details are read live from your traces.
- Only agents that already have a diagram in the window appear in the agent picker.
What your traces need
| To see | Send |
|---|---|
| The agent | Traces with an agent ID. With the ANTS Platform SDKs this comes from the agent name. Traces without an agent ID are skipped. |
| Models | Generation observations with a model name. A generation without a model is shown as Unattributed. |
| Tools | Named spans under the agent |
| MCP servers | Tool spans named mcp__<server>__<tool>. The box is the MCP server, not the individual tool. |
| Sub-agents and Internal steps | Nested spans. A span that calls a model is a sub-agent; a span with only child spans is an internal step. |
| Datastores and External services | For SDK pipelines: span metadata step set to database for a datastore, or service set to a service other than the agent's own for an external service. An optional operation value such as select or insert marks the call as a read or a write. |
| Failures | Observations at level ERROR, with a status message to explain the failure |
| Session filter | A session ID on traces |
ANTS Platform recognizes Claude Code, Codex, GitHub Copilot, CrewAI crews and SDK pipelines, and has verified its rules against real recordings of each. For other sources, such as Bedrock and Vertex agents or generic OpenTelemetry, the page shows Coverage not yet verified. Anything the rules cannot classify is counted as Unattributed rather than dropped.
Open an agent's map
Pick an agent
Select Select an agent and search by name. The picker lists the newest 50 agents; type to find more.
Optionally filter by what it uses
Type a tool, server or model in the filter, for example jira, github or gpt-4o, and press Enter to pin it. With no agent selected, the picker then lists only agents that used everything you pinned. You can pin up to 8 terms.
Choose a view
Switch between Topology, Dependency and Data-Flow.
Without an agent selected, the page lists agents that had a failed call in the last 7 days, with the target they fail on most. Select Investigate to open the failure details.
Topology
What the agent touched. The agent is on the left and every model, tool and service it reached is grouped by kind. A line means one called the other, labelled with the kind of operation: inference, read, write, exec or internal. Each box shows its calls and failures.
Box borders show the boundary:
- Inside: your own steps, tools and datastores.
- Model vendor: models.
- External service: MCP servers, external services, and well-known services such as GitHub, GitLab, Bitbucket, Jira, Confluence, Slack, Notion, Linear and web search providers.
Use the kind chips (Sub-agents, Models, MCP servers, External services, Datastores, Tools, Internal steps, Unattributed) to show or hide groups. When a map is very large, only the busiest boxes are drawn and the header shows how many more were left out.
Dependency
What the agent relies on. Targets are grouped by kind and ranked by how many traces used them at least once, not by call count. Each target shows "used in X of Y traces". Something used in every trace is a dependency; something called many times inside one trace is a loop. Targets where at least 10% of calls fail are marked with their failure rate.
Data-Flow
One real trace, replayed step by step in the order it happened. The left side is your organization; the right side is model vendors and external services. Each time the flow crosses to the right, data left the organization:
- prompt out and answer back for model vendors
- request out and response back for external services
ANTS Platform does not read the content of these calls, so a crossing tells you that data left, not what it contained.
To choose the trace, use Filter by day, Filter by session and Errors only, then pick from the trace list. The newest matching trace is selected by default. Play trace steps through the replay at 0.5x to 8x speed and pauses on the first failed step. Select any step to open that span in the trace view.
Targets table
Under every view, a table lists each target the agent reached in the last 7 days with its Target, Kind, Boundary, Calls, OK, Failed, Traces and What we know:
- observed directly: the trace recorded this call.
- derived from structure: inferred from how spans nest, for example a sub-agent.
- cannot see inside: the call is opaque, for example a shell command.
A blank count means nothing was observed, not zero.
Investigate failures
Select the failure marker on a box to open the failure details for that target: how many calls failed, the error messages grouped by text with a hint (for example timeouts, authentication or rate limiting), and each failed call with Replay this trace and Open span.
Limits
- The window is fixed at the last 7 days.
- An agent-day with an unusually large number of spans is not rebuilt, and the previous diagram is kept. A single trace that is too large is left out, and the header shows Collection gap in the last 7 days.
- A Data-Flow replay draws the first 60 steps, with Show N more steps for the rest.
Planned capabilities are listed on the roadmap.
Next steps
- AI Agent Center: cost, latency, risk and compliance per agent.
- AI Inventory: the agents, models and tools in use.
- Traces: instrument your agents so their traces carry the detail the map needs.