AI agents for project management are software teammates that read your projects, do routine work like triage, digests and status updates, and propose bigger changes for a person to approve. In AlianHub, every agent has a skill, a list of allowed actions and an autonomy level you choose. The risky actions simply do not exist for agents, so you get the speed without handing over the keys.
What an AI agent is in AlianHub
An agent is not a chat window. It is a member of your workspace with an owner, a run budget and a scope. An AGENT tag marks it everywhere it appears, so nobody mistakes its comments for a colleague's.
Each agent is built from three parts:
- A skill, such as a digest, a brief parser or a bug triage playbook.
- Allowed actions, for example comment, create a subtask or draft a doc. Nothing outside that list can run.
- One autonomy level, from L0 to L3, which decides how much it may do without asking.
Only the model is external. The skills, the actions and the limits live in AlianHub. You can pin a model per agent or follow the workspace routing policy.
You assign work to an agent the way you would to a person. And an assignment can always return to a person with the work preserved: an agent can decline before starting, hand back partway with its notes and hours logged, or be paused if it runs too long or goes off track.
The autonomy ladder: four levels of freedom
The core control is the autonomy ladder. You give each agent exactly as much freedom as the job needs.
| Level | What the agent does | What waits for a person |
|---|---|---|
| L0 | Reads, answers and suggests | Every change goes to the AI Inbox |
| L1 | Proposes every change | Nothing is written until approved |
| L2 | Runs low-risk, reversible changes | Everything else waits |
| L3 | Like L2, and may also run on a schedule | Everything that is not low risk |
Low risk means things like a comment, a nudge, a status change within a list, or a report. A project can also be stricter than the agent, so a sensitive project can hold every agent to proposals only.
The AI Inbox
When an agent wants to do something beyond its level, the proposal lands in the AI Inbox. A person approves or declines it. Moving tasks between lists, assigning work and similar wider changes wait there by default.
Undo and the audit log
Changes an agent makes on its own come with Undo. Every action, whether the agent acted, proposed or was refused, is written to the audit log. Each agent also keeps a revision history of every settings change.
What agents can never do
This is the part that matters most for trust. Some actions are not switches you might forget to turn off. They do not exist as actions for agents, so even a stolen token cannot reach them:
- Delete a project or a task
- Anything to do with billing
- Deploy to production or merge code
- Remove a member or change permissions
- Mark a task Done
That last one is deliberate. No agent closes a task in any project. A person checks before Done, whatever the project has chosen. Attempts at hard-stop actions are logged.
For code teams, CI keeps running where it runs today. AlianHub reads the result from GitHub or GitLab and never builds or deploys.
Tip: There is one switch to pause every connected agent across the whole workspace. Know where it is before you turn on your first agent. A change a person already approved is still made.
Budgets and limits
Agents cost money when they call a model, so AlianHub puts limits on runs:
- A daily action limit per agent, shown as actions used today against the cap.
- A monthly spend cap, shown as spend this month against the cap.
- Run now, Pause and Stop controls on each agent.
You also choose where an agent runs. A workspace key runs agents server-side, paid by the company and capped per agent, which suits scheduled work such as a daily digest. A personal AI account means a person uses their own Claude Code subscription, with no workspace spend. A self-hosted model, such as Ollama or vLLM on your own GPU, keeps everything inside your network, which suits regulated teams and air-gapped installs.
Real use cases for project management agents
Here is how teams put agents to work, based on the skills and team packs AlianHub ships.
Weekly digest for leadership
A Reporter agent at L3 runs on a schedule, reads the open tasks in a project and posts a digest of what moved and what is stuck. If it thinks two stuck tasks should move to Blocked, that move waits in the AI Inbox.
Intake and request sorting
An intake agent parses incoming briefs and sorts new requests. With the router you can select many tasks at once; it proposes an agent per task and refuses the ones that need a person.
Bug triage
A Bug Triager comments on new bugs and keeps the task current. It may set In progress or In review, attach a link to the pull request and propose publishing to a test site. Production deploys are a human-only hard stop.
Team-specific role agents
Team packs turn on a team's role agents in one click, across industries such as IT companies, agencies, manufacturing, e-commerce, construction, education and clinic administration. Examples include a Support Agent, an Invoice Preparer and a Snag List Keeper. All start paused, and a pack can be undone. As a rule of thumb, plan one person to supervise three to five agent seats.
Every new workspace also gets a Workspace Guide you can @mention on a task to ask what comes next. And the AI project generator turns a brief into a full plan of lists, tasks and views that you review before it is built.
How to start with AI agents safely
A sensible rollout looks like this:
- Start at L0 or L1. Let the agent answer and propose for a week or two. Read its proposals in the AI Inbox.
- Pick one boring, repetitive job. A weekly digest or request sorting is ideal: low risk and easy to judge.
- Set a spend cap and a daily action limit before the first run.
- Make sensitive projects stricter than the agent, so they stay proposal-only.
- Review health and quality. AlianHub shows error rate, approval rate and cost per agent and per model, plus a quality view over 7, 30 or 90 days.
- Promote slowly. Move an agent to L2 only when its proposals are consistently approved, and to L3 only when you want it on a schedule.
You can see all of this on the AI page and in more detail on the AI controls page. If your team already uses Claude or ChatGPT, you can also connect your own AI so it works inside AlianHub as you.
FAQ
Can an AI agent delete my projects or tasks?
No. Deleting a project or a task is not an available action for any agent. It does not exist in the action registry, so it cannot be switched on by mistake.
Do AI agents mark tasks as complete?
No. An agent may set a task to In progress or In review, but a person always closes it.
How do I control what agents cost?
Each agent has a daily action limit and a monthly spend cap. You can also run agents through a personal AI account or a self-hosted model, which adds no per-run workspace spend.
Can I try AI agents before installing anything?
Yes. The live demo lets you explore the AI features in a demo workspace. Demo data may reset periodically.
Get started
AlianHub is open source, self-hosted and has no per-seat pricing. Open the live demo, take the product tour, and turn on your first agent at L0 to see what it would do before it does anything.