Agent Pool

The Agent Pool is a Polarion project that holds everything a team can hire: the agents, the skills they follow and the connections to other systems they may use. Project leads hire agents from the pool into their own work projects. Nothing in a work project changes when you edit the pool, except that the next run of a hired agent uses the updated definition.

The pool is any empty Polarion project that you set up with Administration → Setup (see Installation). The Agent Runner finds it by itself: it is the project whose Agentic settings item has the role pool. There is one pool per server.

The Agency page

The pool has two pages in its navigation: Agency is the catalogue the people who define agents work with, Administration is the operator’s side. Open the pool project and choose Agency. It has four tabs:

Tab	What you see
Agents	Every agent with its model, skills, connected systems, the projects it works in (load against capacity) and its spend so far.
MCP servers	The registered connections, their tool policy at a glance and tools held for review. Register MCP server starts here.
Skills	The skill library with status, number of tools, the agents that have each skill and when it last changed.
Bullpen	The catalog of ready-made skills, agents and connections that ships with the release, compared with the pool. Hire what you need as drafts. See Agent Bullpen.

The Administration page

Administration holds everything an Agent Pool administrator does to keep the platform running. It has four tabs:

Tab	What you see
Setup	Agentic setup: the pool’s configuration (version, Agent Runner address, upgrade and repair), the AI-enabled projects — enable, check and upgrade them — and the search box to enable another project. See Installation and Enable a project.
Credentials	The model key, one Polarion token per agent and the credentials of connected systems: set, replace and test them; missing items are red. See Credentials.
Gateway	Agent actions in connected systems over the last days, per server, blocked attempts and the standing rules people have set. See Governance.
Sandboxes	The sandboxes agents work in right now, with resources, allowed sites, commands and files. See Sandboxes.

In a brand-new pool only Agency is in the navigation at first. When Agency opens and the Administration entry is missing, it shows a banner The Administration page is not in this pool’s navigation yet with the button Add Administration to the navigation; one click adds the entry, as you, and Administration → Setup takes it from there (see Installation).

One Polarion user per agent

Every agent acts as its own Polarion user, its bot user. Because of that:

• Work items, comments and links an agent creates show the agent as their author in Polarion’s history.
• An agent is assigned work like a person: you set it as the Assignee of a task.
• An agent only sees projects it is a member of. A bot that is not hired into a project gets “not found” for it.

To create a bot user:

1. Create a Polarion user whose id starts with bot-, for example bot-atlas. The hire and let-go actions only accept user ids of the form bot-<lowercase letters, digits, dashes>, and the runner treats comments from bot- users as agent comments.
2. Assign it a Polarion licence and the global role user.
3. Create a personal access token for it and paste it into Administration → Credentials (the agent’s row links to the token page in Polarion). The agent shows as missing there until you do. See Credentials.

Caution

Don’t give bot users project roles by hand. Hiring adds the roles project_user and project_assignable to the work project, and letting the agent go removes them. Both go through Polarion’s security service, so Polarion also keeps its repository access rules up to date.

Agent fields

An agent is a work item of type Agent in the pool.

Field	Meaning
Title	The agent’s name, shown everywhere (e.g. “Atlas”).
Description	The persona or job description. It becomes part of the agent’s instructions.
Role	Job title, e.g. “Requirements Engineer”.
Bot user	The Polarion user the agent acts as.
Owner	The person responsible for this agent definition.
Model	<provider>:<model>, for example anthropic:claude-sonnet-5-5. Providers: anthropic, openai, google.
Harness	Where the agent works: Built-in (in the runner), Claude Code or Codex (each in its own sandbox per task).
Capacity	How many engagements the agent may have at the same time, across all projects.
Avatar	An emoji used in the views.

Links from the agent:

• has skill → the skills it follows. Only approved skills count.
• uses → the MCP connections it is able to work with. A project still has to grant each one when hiring.

Note

The Claude Code harness works with Anthropic models and the Codex harness with OpenAI models. If the runner has no sandbox service configured, sandbox agents run in the built-in loop instead. See Sandboxes.

Agent lifecycle

stateDiagram-v2
  direction LR
  [*] --> Draft
  Draft --> Available: publish
  Available --> Retired: retire
  Retired --> Available: reinstate
  Available --> Draft: rework
  classDef ok fill:#e3f2e7,stroke:#1f7a3d,color:#143d22
  classDef mute fill:#f1f4f6,stroke:#8a96a3,color:#2a3440
  class Available ok
  class Draft,Retired mute

Agents have a Polarion workflow, so publishing an agent is a deliberate, recorded step.

Status	Meaning	Actions
Draft	Being defined. Can’t be hired.	Publish to the pool → Available
Available	Can be hired and gets work.	Retire → Retired, Rework → Draft
Retired	Out of service. Can’t be hired.	Reinstate → Available

Only Available agents appear in the hire wizard, and only while they are below their capacity and not already hired into that project. Before you retire or rework an agent, let it go from the projects it works in (Team page) so that no open tasks still wait for it.

What a project gets when it hires

Hiring is done by a project lead on the project’s Agent Team page (see the user guide). The wizard creates an Engagement work item in the work project, linked to the pool agent, with:

• the supervisor, who answers the agent’s questions and approves its results,
• a budget in USD (see Budgets),
• optional project instructions that are added to every task,
• the granted connections, each with an optional scope and notes,
• for sandbox agents, the internet sites the agent may read.

The person who hires needs the right to manage the project’s members, because Polarion checks their permission when the bot user is added.

Filling a pool

The way to fill a pool is the Agent Bullpen (Agency → Bullpen): the catalog of ready-made skills, agents and connections that ships with every release. You tick what you need and it arrives in the pool as drafts, created as you; Approve and publish then walks the skills and agents through their reviews, and connections become active once an administrator has set their credential and reviewed their discovered tools. The pool’s configuration itself (types, fields, workflows, forms) comes from Agentic setup, not from the catalog.

You can also create agents and skills by hand in Polarion, like any other work item. Remember to publish agents and to have skills reviewed before you expect them to work.

For test setups the runner still ships a batch alternative: scripts/seed_pool.py creates a sample set of skills and agents and walks them through their workflows. It is idempotent and matches items by title.

cd runner
uv run python scripts/seed_pool.py

Related

• Agent Bullpen, Skills and Connections
• Governance for the full oversight model
• Concepts and Architecture