AI agents

Agents are AI teammates that work on your issues. Give one an issue and it reads the issue and its discussion, works in your repository, pushes a branch, and Deckbell opens the pull request. Agents never take a seat.

Roles

Every agent has a role that decides what it does:

Role	What it does	Changes code
Coder	Implements features and fixes, then a pull request is opened	Yes
Reviewer	Reviews a pull request for correctness and quality	No
Security	Audits changes for secrets, injection, auth and dependency risks	No
Tester	Writes and fixes automated tests, then a pull request is opened	Yes
Triage	Investigates an issue: likely cause, what's affected, a plan and a size	No

Coder, Reviewer and Security are on every plan; Tester and Triage come with Pro. On Business you can also write custom roles: open Roles on a project's Agents tab, choose New role, and give it a name, instructions and whether it's read-only. Custom roles are shared across your workspace. A role doesn't choose a model: agents on Deckbell use Auto, and agents on your own key use the model picked on the agent.

Creating an agent

Use New agent on a project's Agents tab. The dialog walks you through:

1. How it runs:
   • Deckbell (recommended): you don't pick a model. Deckbell chooses the right Claude model for each step (see Auto below) and the run uses your workspace's AI credits.
   • Your own Anthropic key: paste your API key (stored encrypted) and pick one model for every run; Anthropic bills you directly. See Choosing a model on your own key.
2. Details: a name (people @mention the agent by it) and a role.
3. Repositories: the GitHub repositories the agent can work in. Connect GitHub first in Integrations.

Both options show the estimated cost per typical task, so you can compare them before you choose. These are estimates for a typical task; real runs cost more or less depending on the work.

How Deckbell picks models (Auto)

Agents that run on Deckbell use Auto: Deckbell codes with Claude Sonnet 5.5, saves Claude Opus 5.5 for hard tasks, and escalates when the agent gets stuck.

1. Size the task. A quick check reads the issue's title, description, labels, estimate, sub-issues, linked issues and how much discussion it has, and calls it small, normal or hard (when in doubt, normal).
2. Small tasks go straight to coding on Claude Sonnet 5.5.
3. Normal and hard tasks get three steps:
   • Plan: the code is read and a concrete plan written, without changing anything (Claude Sonnet 5.5 for normal tasks, Claude Opus 5.5 for hard ones).
   • Code: Claude Sonnet 5.5 follows the plan, pushes the branch, and Deckbell opens the pull request.
   • Review: the change is reviewed and the review posted on the pull request (Claude Sonnet 5.5 for normal tasks, Claude Opus 5.5 for hard ones).
4. Escalation: if the coding step fails or pushes no changes, Deckbell retries it once on Claude Opus 5.5.

Small tasks are reviewed too when you turn on Review small tasks too for the agent (the review then uses Claude Sonnet 5.5). Read-only roles run in one step: pull request reviews use Claude Opus 5.5, other read-only work (such as Triage) uses Claude Sonnet 5.5.

Each run has a cost limit. Before each step Deckbell checks what the run has spent; if the next step would take it past the limit, the run stops with a message saying which step was skipped and why. A pull request that was already opened stays open.

When you chat with an agent that runs on Deckbell, short questions are answered by Claude Haiku 5.5 and longer requests by Claude Sonnet 5.5.

Choosing a model on your own key

On your own key you pick the model, and every run uses it (no routing). The dialog compares the current models side by side:

Model	Best for
Claude Opus 5.5	Hard, ambiguous work: architecture, tricky bugs, large refactors
Claude Sonnet 5.5 (default)	Everyday coding: features, fixes and tests
Claude Haiku 5.5	Small, well-defined chores: copy changes, triage, simple scripts

Each card shows a relative quality bar, speed and the estimated cost per typical task on your key. Below the cards, a note compares this with Deckbell Auto, which uses Opus only where it pays off (hard tasks) and costs less per typical task than Opus on your key (by our estimate).

Agents created on older models (such as Claude Sonnet 5 or Claude Opus 5) keep working and are marked older model; pick a current model to switch.

Editing an agent

Choose Edit in an agent's ⋯ menu to change its name, role, how it runs (and, on your own key, its model), repositories, base branch or auto-review. You can switch an agent between Deckbell and your own key at any time: switching to your key asks for it once (leave the field empty later to keep the saved key) and starts on Claude Sonnet 5.5; switching to Deckbell removes the saved key and moves the agent to Auto. The change applies from the next run.

Starting a run

• Assign an issue to the agent.
• @mention it in an issue comment: "@Coder can you take this?"
• @mention it in a channel with an issue key: "@Coder pick up HRB-142", or in an issue's thread in the board's channel.
• Press Run on the Agents tab.

The agent answers where it was asked: "On it. Working in harbor/billing on branch coder/hrb-142." In a channel it replies in the channel (or in the thread, if you asked in a thread), and the result is also posted on the issue. Thanking an agent ("thanks @Coder") doesn't start a run.

What the agent knows

Each run starts with the issue's title, description, status, priority and labels, its parent and sub-issues, linked issues (blocks, blocked by, related) with a short excerpt, the last 30 comments, and what you asked when you @mentioned it.

Results and pull requests

A coding agent works on one branch per issue (for example coder/hrb-142); running it again on the same issue continues that branch and pull request. When it finishes, Deckbell opens the pull request and posts a PR card on the issue (and in the channel when asked there) with the lines added and removed, files changed, how long the run took and what it cost. Open the run from the card to see every step it took.

To see everything an agent has done, open Runs on its card on the Agents tab. It lists every run, newest first, with its status, the issue (or a short summary), what started it and who, the pull request, how long it took and what it cost. Choose Load more to go further back, and choose a run to open its page.

The run page also has a Steps table: each step (size, plan, code, retry, review, or a single run on your own key), the model it used, its input, cached and output tokens, and its cost. The run's cost is the sum of its steps.

If the agent didn't push any changes there's no pull request, and the result says so. When a run can't start or fails, the agent says why.

Automatic pull request reviews

Reviewer, Security and other read-only agents can review every new pull request in their repositories. Turn on Review new pull requests automatically when creating or editing the agent. When a pull request is opened, reopened or marked ready for review (drafts and pull requests from forks are skipped), the agent reads the change and posts its findings as a comment on the pull request and on the linked issue, or in the board's channel when no issue is linked. This needs the Deckbell GitHub App.

Pull requests are linked to an issue when their branch, title or description contains its key, for example HRB-142. Linked pull requests show on the issue with their state.

Chatting with an agent

You can also chat with an agent directly from the Chat area.