See which AI agent starts which
The relationship mindmap in ThoughtShieldAI shows, for any AI agent you pick, the agents it starts and the agents that start it, one level at a time. It is built from the logs your agents already write: no SDK, no code change, and the data stays on your machine.
A sample fleet: double-click an agent, unfold who it starts, hover for the data passed along each arrow, fold back.
Why agent lineage matters
AI agents rarely work alone. An agent that processes invoices may hand line-item checks to a second agent, which hands purchase-order matching to a third, which calls a fourth to read the finance system. Each hand-off costs money, moves data and can fail, and none of it is visible from outside.
ThoughtShieldAI records every hand-off as a spawn: one run of an agent starting a run of another. The fleet view draws all of them at once, which answers "what do we run?". The mindmap answers the narrower question you usually have: what does this agent touch?
- Which agents does invoice-processing-agent start, and how often?
- If a subagent fails, which workflows depend on it?
- Where does a chain of delegation end?
- How much text passes along each hand-off, which matters for cost and for data exposure?
How to use it
- Open the ThoughtShieldAI dashboard and choose Lineage. The relationship graph shows the whole fleet.
- Double-click an agent to grow a map from it. Agents it starts appear to its right; agents that start it appear to its left.
- Tap a ⊕ badge (its number says how many agents are hidden) to unfold the next level. Repeat to follow a chain as far as it goes.
- Hover an agent to see its runs and, on each arrow, how much text went down to the subagent and came back.
- Tap − to fold a branch, Fold all to start over from the root, or ← All agents to return to the fleet.
What the map shows
Every number on the screenshot matches a row in the table below it.
| # | What it is |
|---|---|
| 1 | The root: the agent you double-clicked. |
| 2 | An agent the root starts, placed to its right. |
| 3 | ⊕ badge: this agent starts 2 more agents not shown yet. Tap to unfold them. |
| 4 | − badge: this agent is unfolded. Tap to fold its branch away. |
| 5 | Spawn count: the root started this agent 28 times. |
| 6 | A third level: the chain continues as deep as the data goes. |
| 7 | Where you are: 3 agents unfolded, 7 shown, 21 hidden. ← All agents returns to the fleet. |
| 8 | Fold all keeps only the root's direct relationships. |
| 9 | The summary line: the root, how many agents and relationships are on the map. |
What it does not do yet
- One unfold shows at most 12 agents, the busiest first. The details drawer always lists every relationship.
- The map is not saved. It resets when you reload the page or leave the Lineage view.
- The canvas is not yet readable by screen readers. The details drawer is the text alternative for now.
Under the hood
Relationships are not stored; they are folded at read time from run-level spawn records, so the map always reflects every execution ThoughtShieldAI has seen. The whole map is two values, a root and a set of unfolded agents, recomputed by one breadth-first pass in your browser. Everything is read-only.
The full design document covers the data model, the algorithms, the layout engine and how it was tested, with code.
Questions
Does using the map change or slow down any agent?
No. ThoughtShieldAI only reads the logs and records agents already produce, and the map is computed in your browser from data already collected.
Which platforms does it cover?
Agents discovered from Claude Code, Codex, OpenClaw and Hermes on the machine, plus catalog-level inventory from Google Vertex AI, Amazon Bedrock, Cloudflare and Azure AI Foundry. Spawn relationships come from the local agent logs.
Why does an agent appear only once on the map?
The map is a tree, so each agent sits under the first branch that reaches it. If a second branch also leads to it, that link is still drawn as a line.
What do the numbers on the arrows mean?
How many times the agent on the left started the agent on the right, across all recorded runs.
Is it open source?
Yes. The core is MIT licensed and self-hosted: it stores data in a local SQLite file and the dashboard binds to your own machine.
Run it on your own fleet
One clone, one scan, one dashboard. No API keys pasted and no change to your agents.
git clone https://github.com/ANVEAI/anveinspect
cd anveinspect && npm install
npx anveinspect doctor
npx anveinspect dash