AI agents for the rest of the team: PMs, QA and clients

A PM and an advisor agent write a task description together; the agent asks a question and the artifact is posted.

Not everyone on the team writes code

Most AI agent tools are built for developers: they read the repository, edit files and run tests. A large part of a product team never does any of that.

A project manager turns a request into a clear task. QA writes test cases. An analyst or a designer writes requirements or a brief. And the client, whose idea it was, knows answers nobody else has. Their work is text, and it lives in the task tracker.

CRUX has two kinds of agents for this reason:

  • Worker agents run on a developer's machine, in Claude Code, and write code.
  • Advisor agents run on our server, as chats in the CRUX dashboard. They are for people who mostly work with the text of tasks.

Advisors work for clients too. A client can answer an advisor's clarifying questions right in the task, and the answers are recorded there as an artifact, with no PM relaying them through calls and emails. We wrote about that separately: Bring your client into your tasks: answers straight from the source.

This article is about advisors. New to CRUX? Start with Stop writing prompts: why AI coding agents need a workflow.

Read-only, on purpose

An advisor can read the project's wiki and code. It can open files, search them and look at their history, but it cannot change them. There is no "write" level for code or the wiki, and there is no shell to run commands. You also decide per agent whether it sees the code, the wiki, both or neither.

What an advisor can do is work with the task, and only in the ways its job needs:

  • Most advisors read the task and post one artifact: a comment with the result of their step, such as a clarification, requirements or test cases.
  • Some get one or two more tracker actions. An estimator can set an estimate. A project-manager agent can create subtasks. A bug-intake agent can create a bug and attach a screenshot.
  • In our templates, nothing goes to the task until you approve it. The agent shows a draft in the chat, you correct it, and it posts only after an explicit "ok". Anything destructive, like deleting, either asks you to confirm with a click or is refused.
  • Agents don't move tasks between statuses. A person does that.

Chats themselves are kept for a limited time that you can set, then deleted. The approved artifact on the task is what stays.

Understand a project you didn't write

You don't need to know how the project is built to find your way in it. Maybe you are not a developer, or you joined last week. An advisor with access to the wiki and the code reads them for you and explains, in plain words, what exists, how a feature works today and what was decided.

A new QA engineer might ask: "How does password reset work right now, and what happens if the link has expired?" The answer is a short, plain description: the steps the user goes through, how long the link stays valid, the message they see when it has expired. It names the wiki page and the code file it came from, and it says so when the two disagree. The agent only reads and explains.

A small toolbox, on purpose

We could give every advisor every tool we have. We don't.

When an agent sees forty tools, it spends effort choosing and sometimes picks the wrong one: it searches when it should read, or edits when it should comment. With a short list, the right choice is usually obvious.

So an advisor gets a small, fixed set: read a file, search, look at history, follow wiki links, search the web, load a skill. On top come only the tracker actions its job needs.

Its own skills and MCP servers

Each advisor can have its own set of skills and MCP servers.

  • A skill is a written method the agent follows: how to run an interview, how to write test cases, how to estimate. It is a folder with a SKILL.md file in your repository. You can write your own or add a public one from GitHub.
  • An MCP server connects the agent to another system: a design tool for the designer agent, your analytics for the analyst, your documentation site for QA.

An agent sees only what is attached to it.

Example: a PM, a vague task and grill-me

grill-me is a public skill by Matt Pocock. It interviews you about an idea until you both share the same understanding, and it suggests an answer for each question so you can just say "yes". It is a thin wrapper around a skill called grilling, so you add both. By default it asks in short rounds; one line in the agent's prompt makes it ask one question at a time, which suits a quick chat.

Attach it to a PM advisor and open a task that says only: "Let customers export reports. Several asked for it."

Agent: I read the task and the wiki. The wiki lists three reports: Sales summary, Activity log and Invoices. The three customer requests linked in the task all mention Sales summary. Suggestion: start with Sales summary only. Right?

PM: Yes, Sales summary only for now.

Agent: Which format? The code already has a CSV download for Invoices that we could reuse. Suggestion: CSV first, since it opens in Excel; a real Excel file later if customers still ask. OK?

PM: They said "Excel", but CSV is fine if it opens there.

Agent: Who may export? According to the wiki, admins and managers can see Sales summary. Suggestion: whoever can see the report can export it.

PM: No, admins only. Managers shouldn't take the data out.

Agent: Should the export follow the filters on screen (date range, team), or always include everything? Suggestion: follow the filters.

PM: Follow the filters, and cap the range at 12 months.

Agent: Here is the draft for the task. Should I post it?

PM: ok

Agent: Posted to the task.

The artifact on the task:

ARTIFACT: clarification v1 (approved by the PM in chat)
- Scope: export of the Sales summary report only; other reports are out of scope.
- Format: CSV (opens in Excel), reusing the existing CSV download.
- Access: admins only; managers can view the report but not export it.
- Data: follows the on-screen filters; date range capped at 12 months.
- Open: should exports be logged for audit? Not answered yet.

Ten minutes of chat, and the next person (a developer, QA or another agent) starts from a clear task instead of a guess. (How we keep that wiki current is in Our project wiki: memory that AI agents can read and write.)

No commission on tokens

Each project connects its own key from a model provider, so you pay the provider directly, at their price. We don't resell tokens and we take no commission on them.

In short

  • Advisors are for people who work in tasks, not code: PMs, QA, clients.
  • They read the wiki and code for you, and never change them.
  • A small set of tools, plus each agent's own skills and MCP servers.
  • They draft, you approve, then the result goes to the task.
All articles