Agents

Understand Cloud Agents system agents and custom agents, how to design them, and how to use them for reliable repository work.

In Cloud Agents, an agent is a reusable operator definition. It combines role, model, instructions, and workspace constraints so the same class of work behaves consistently from run to run.

Agent Types

Type	What it is	Best for
System agent	Built-in Cloud Agents operator maintained by Skytells	Default review and execute workflows
Custom agent	Workspace-specific agent created by your team	Specialized repository tasks or team conventions

The settings view in Cloud Agents shows both built-in agents and custom agents in one place so operators can see what is available, which agents are enabled, and how many custom-agent slots are currently used.

Built-In Agents

Common system agents visible in the workspace include:

Eve Agent for broader orchestration-aware execution flows
Execute Agent for change-oriented workflows
Code Review Agent for review-first workflows

System agents are the right starting point because they align with the product's default safety and execution model.

What a Good Agent Definition Includes

Every useful agent definition answers five questions clearly:

Question	Example answer
What is this agent for?	Review pull requests for reliability and test gaps
What can it touch?	UI routes under `src/app/approvals/**` only
What should it avoid?	Billing, auth, and infrastructure code
What standard should it follow?	Existing design system and current test conventions
What should it return?	Findings, precise file targets, and validation notes

When to Create a Custom Agent

Create a custom agent when the same class of work repeats often enough that it deserves its own operator contract.

Examples:

a migration agent for framework upgrades
a docs sync agent for keeping docs aligned with implementation changes
a release hardening agent for pre-merge validation
a security triage agent for dependency or path-sensitive review

Do not create a custom agent simply because a task is important. Create one because the task pattern is stable.

Creating a Custom Agent

Open Settings → Agents

Go to Settings in Cloud Agents and open the Agents area.

Select New agent

Use New agent to create a custom agent profile.

Name the role, not the model

Prefer names such as Docs Maintainer, Payments Reviewer, or Frontend Execute Agent over names that only describe the backing model.

Write instructions for repeatability

Define what the agent should optimize for, what it must not touch, and what output form it should prefer.

Choose the model deliberately

Pick the model that fits the role. Heavy review and focused code execution often benefit from different model characteristics. See Model Catalog.

Enable and test on a low-risk repository

Turn the agent on and validate it on a constrained repository before broadening use.

Prompt Design for Custom Agents

Good instruction traits

Specific scope

The agent knows what folders, files, or pull-request classes belong to it.

Clear prohibitions

The agent is told what it must not modify, approve, or ignore.

Stable output shape

The team knows whether to expect findings, a plan, a change set, validation notes, or all of the above.

Weak instruction traits

broad goals with no file or domain scope
conflicting objectives such as "move fast" and "touch everything carefully"
missing validation expectations
no statement about prohibited paths or sensitive systems

Agent Governance

Custom agents should be managed like any other operational control:

keep the number of active custom agents small enough that operators remember what each one is for
disable agents that no longer have a clear owner or use case
review instructions after repo structure changes, framework upgrades, or team-process changes

Choosing the Backing Model

Cloud Agents lets you align the agent role with the model that powers it. The model choice should reflect the work:

review agents benefit from strong reasoning and consistency
execute agents benefit from precise code editing behavior
specialist agents may benefit from domain-specific instruction tuning or a larger context window

See Model Catalog for currently available models and Eve for orchestration context.

If a team is new to Cloud Agents, start with system agents first, then create a custom agent only after you can name a repeated failure or repeated manual step that the custom agent would remove.