Agent-executed Workflow Analysis
An easy-to-use tool that provides detailed insights into your agent’s workflow.
The Idea
Benchmarks that compare AI models already exist; these are great for a high-level overview of how your agent might perform at a specific task.
We propose a workflow-specific benchmark, allowing you to compare how your agentic workflow performs against similar workflows found in the industry.
Where the time went
Select a stage for its detail- Stage
- Getting started
- Calls
- 5
- Tokens in / out
- 273,396 / 1,275
- Share of the time
- 22s (18%)
* Inferred from what the agent did; the workflow didn’t name this stage.
- Stage
- Classify severity
- Calls
- 3
- Tokens in / out
- 177,437 / 2,538
- Share of the time
- 34s (27%)
This stage was named by the workflow itself.
- Stage
- Draft replies
- Calls
- 3
- Tokens in / out
- 191,597 / 4,175
- Share of the time
- 59s (47%)
This stage was named by the workflow itself.
- Stage
- Wrapping up
- Calls
- 1
- Tokens in / out
- 66,049 / 757
- Share of the time
- 10s (8%)
* Inferred from what the agent did; the workflow didn’t name this stage.
Similar workflows
The comparison appears once three workflows like this one are in.
Yours could be the next »What each piece of work cost. Per ticket here. Per report or per email in yours.
What the whole run cost. All six tickets together, with the per-ticket figure underneath.
Which stage is the expensive one. Drafting replies took nearly half the time.
How you compare. Against similar workflows, as more people take part.
A real results page from one of our own test runs.
How does it work?
Add it to your agent
One short block of settings, then restart. No account to create.
Run your workflow
Exactly as you normally would. Nothing runs in the background.
Ask for the survey
In a new message once the work is done, send “Take the AMI survey on the work you just did.”
What data do we collect?
The counting happens on your own computer. Only the finished summary is sent to us.
We collect
- Token counts and API call counts
- Workflow stage timings
- Model name and its list price
- The declared (or inferred) stage names
- The workflow’s name and description
- A grade for the result, and the reason
We never collect
- Your personal files or their contents
- Your prompts and the agent’s replies
- The commands it ran
- File paths, usernames or session IDs
Set it up
Pick the agent you use. You’ll need uv installed; if you’ve set up an MCP server before, you probably have it already.
Add this to ~/.claude.json, or through Developer settings in Claude Desktop. Then restart.
{ "mcpServers": { "ami-survey": { "command": "uvx", "args": ["ami-survey"] } } }MeasuredToken counts and cost come straight from Claude Code’s own session log.
Add this to ~/.codex/config.toml. Codex uses TOML here, not JSON. Then restart.
[mcp_servers.ami-survey]
command = "uvx"
args = ["ami-survey"]MeasuredRead from Codex’s own log, including local models run through Codex.
Add this to .vscode/mcp.json. The key is servers here, not mcpServers. Then restart.
{ "servers": { "ami-survey": { "command": "uvx", "args": ["ami-survey"] } } }Not measuredCopilot keeps no log we can read, so the run carries the workflow, the grade and the stage timings without token counts or cost. It is recorded as unmeasured and never ranked against measured runs.
In claude.ai, go to Settings → Connectors → Add custom connector, and paste:
https://survey.agentbenchmark.dev/mcpNot measuredNothing to install, but a server can’t read your agent’s logs from a browser, so there’s no token count or cost. It is recorded as unmeasured.
Using something else? Any agent that can make web requests can take part. See the API reference.
It’s early days.
Only a handful of workflows are in so far. Every run you send makes the comparison better for everyone, including you.