She reads your agent and tells you what is weak

Point her at a repository. She works out whether there is an agent in it, assesses it across nine axes, and tells you what she measured and what she did not.

Create an organization

Works with any public GitHub repository. Paste a link and she reports what she found.

Heros
19things she does
9axes assessed
$1.00ceiling on a run

Every run, before anything is spent

Get started
The plan, written downevery task named up front
10 tasksfor a full assessment
Ceiling $1.00and each task reports its cost
Nothing spent on a refusaland the refusal says why

Nine axes

Nine axes.
Eight of one node, one between them.

model

Which model this node calls, and how it is sampled.

She looks for: a model is named · a sampling parameter is set · a model name appears
prompt

What this node is told, before it is told anything else.

She looks for: a system prompt is defined · a system role message · a prompt template
skills

What skills are bound at this call site.

She looks for: a skill binding · a skill registry
context

What conversation history this node is handed.

She looks for: the message list is built · history is spliced into a call · history is trimmed or summarised
tools

What tools it is offered, and which of them it actually calls.

She looks for: a tool list is passed · a tool is declared · tool calling is configured
memory

What it keeps between calls, and where it keeps it.

She looks for: a memory object · state is persisted between calls · a retrieval store
harness

What it can reach, and what it can spend.

She looks for: a limit or retry policy · error handling around a call · a spend ceiling
loop

How many turns it takes, and in what loop.

She looks for: an iteration ceiling · an unbounded loop · an agent control loop
graph

The order the nodes run in, and who hands off to whom.

She looks for: a graph is constructed · control passes between agents

Eight of these are properties of one node. graph is a property between nodes — which is why it is the one axis that can never be read off a single call site, and the one she is most likely to tell you she could not measure.

What she does

Nineteen things.
That is the list, not a sample.

  1. What does my agent actually do, step by step?
  2. What conversation history does this node get?
  3. What does this node remember between calls?
  4. What tools is it offered, and which does it call?
  5. What skills are bound at this call site?
  6. How many turns does it take, and in what loop?
  7. What can it reach, and what can it spend?
  8. Should these nodes run in this order?
  9. Look at my repository and tell me what is weak.
  10. Build me an evaluation set for this agent.
  11. Is this version better than the last one?
  12. What did you measure, and what did you not?
  13. What happened in that run?
  14. Change what this node is told.
  15. Change which model this node calls.
  16. Change something on an axis and show me the diff.
  17. What exactly would you write into my source?
  18. Fix it, and open a pull request.
  19. Put an approved change into my repository.

Evidence

Every finding points at
a line of your code.

  • file:line Every span she reports carries a path and a line number, and nothing in the code can build one without them. A finding you cannot navigate to is a finding you cannot check.
  • why Each span names the pattern that caught it — a spend ceiling, an unbounded loop — so you can tell a real signal from a word that happened to appear in a comment.
  • absence Finding nothing is a finding, and it is not the same finding as failing to read. One note says “no code matching any memory signal was found”; the other says “the walk hit its limits”. Only one of those is a statement about your agent.
  • the cap Twelve spans an axis, ranked by how close they sit to a call site. Where more matched, the report says how many matched, how many are shown, and how many of those are in a file that calls a model.
  • coverage And it is a question you can just ask her, in those words: what did you measure, and what did you not?

The patterns are deliberately generous, because the two mistakes do not cost the same: a false positive costs you a glance, a false negative costs you a whole axis reported as empty. So what you get back is what was found — not everything that exists.

Who she is

She would rather
refuse than guess.

An abstention costs one exchange. A wrong route spends your money on work you did not ask for — and a conversation is exactly where you are least equipped to notice.

“I cannot route that.”
“Nothing was spent.”
“Give the run a limit.”

Three things she says, verbatim.

Getting started

Three steps.
The first one is free.

01Create an organization
02Paste a GitHub link
03Ask her a question

Create an organization →