---
title: Ask Redline
description: The project agent — it reads your workspace, and it can act in it.
---

Ask Redline is an agent with tools over this project, not a search box. It
decides what to look at, looks, and keeps going until it can answer — and you
watch it happen: every call, its arguments, and what came back.

Ask it a question about your project and it will go and find out, rather than
matching your words against a canned report.

## What it can read

Everything in the project: your experiments and their pass rates, every run and
its status, which agents are connected or online, and your tasks, rubrics,
assets, files and machines.

Crucially, it can open a run in full — the transcript, the final answer, and the
evaluator's note against each criterion. That is the difference between a
dashboard and an answer. *Which runs failed* is a list; *why they failed* is
inside them, and it will go in, read the notes, and check whether the other agent
hit the same thing.

```text title="A question worth asking it"
Why did the last run fail, and what should I change about the task?
```

## What it can change

Four things: it can write a task, write a rubric, launch an experiment, and
cancel one. Anything it creates is linked at the end of the turn, so you can open
it and check the wording before it matters.

> **Launching spends money**
>
> It launches when you have clearly asked for a run in the conversation, and
> then does the whole job in one turn — write the task, write a rubric, choose
> the agents, launch, say what is now running. It does not launch because you
> asked a question about one: *"would that work?"* is not *"run it"*.

It is also told to check what already exists first. An experiment that would
repeat one from ten minutes ago is usually a misunderstanding, and the cheaper
answer is to point at the result you already have.

## How to read the answer

The page is built around the loop rather than around the reply. Its steps appear
as it takes them, grouped under a single line while it works; what each tool
returned is one click away; the answer streams in underneath as it is written.

It is instructed never to state a number, a name, a status or an outcome it has
not read with a tool — and to say plainly when something is missing rather than
filling the gap. An empty project is a fact, not a hole.

## Conversations

Threads are kept. The five most recent are on the opening screen and the rest are
behind **History**.

Reopening one replays it exactly, tool timelines included, because what is stored
is the raw stream rather than a summary of it. Continuing a thread gives the
agent its full context — every earlier tool call and result, not just the prose —
so *"open that run"* and *"now launch it"* resolve against something it read
three turns ago instead of sending it looking again.

## What it thinks with

The same OpenRouter key as the evaluator: the one this Redline was started with,
or one connected in the project. Without a key it says so instead of guessing.