Build an Agent Flamegraph

Agent Flamegraphs group free-form agent activity into stable semantic stacks. Width represents the selected weight, while the stack connects project, agent, intent, tool, process, and effect context.

Choose a tokens, time, operations, files, or network view.

Keep intent tagging reproducible.

Compare related paths without exposing raw prompt text.

Select the question before the view

Use tokens to investigate budget, time to find latency, operations to count behavior, files to inspect repository effects, and network to attribute destinations. Each view projects the same session evidence differently.

Treat semantic labels as analysis inputs

Stable intent labels make sessions comparable and reduce accidental prompt disclosure. Retain the tagging rules or cache used to build a published flamegraph so another reviewer can reproduce it.

Terminal

$ cargo install agentpprof
$ agentpprof --view tokens -o tokens.svg

Primary sources

AgentSight agentpprof guide