Engineer agents

Find the 16 specialist roles that help Engineer Kit explore, plan, build, verify, and deliver software work.

Engineer Kit includes 16 Agents: 10 shared roles used across AgentKit and 6 roles specific to engineering work. Together, they cover the path from an unclear request to a reviewed result without turning every specialist into a separate command you must manage.

How Agents complement Skills

A Skill is a workflow you choose for an outcome, such as planning a change, debugging a failure, or reviewing code. An Agent is a specialist role that receives a bounded delegated task and returns a focused result. A Skill can coordinate one or more Agents, and your coding assistant can delegate to an Agent when the current task fits that role.

Start with the Skill or outcome you need. You usually do not need to select every Agent yourself, and Agent IDs are not a universal slash-command surface. The available discovery and orchestration controls depend on the target runtime.

Find the right help by outcome

The roster below is the complete resolved Engineer set. “Shared” identifies a role included in the common foundation; “Engineer-specific” identifies one of the six roles added by Engineer Kit.

Understand and decide

Agent	Origin	Purpose
`Explore`	Shared	Maps the files, symbols, and relationships relevant to a bounded codebase question.
`researcher`	Shared	Produces a sourced synthesis of technologies, patterns, or primary documentation.
`brainstormer`	Engineer-specific	Explores architectures and solution approaches before implementation begins.
`advisor`	Engineer-specific	Reviews direction at a senior level and returns risks, trade-offs, and recommendations.
`kongming`	Shared	Frames difficult strategic decisions as options, consequences, and a recommended course.

Plan and coordinate

Agent	Origin	Purpose
`planner`	Shared	Turns a significant change into an evidence-based implementation plan.
`project-manager`	Shared	Consolidates work against the plan and reports progress, gaps, and next actions.

Create and improve

Agent	Origin	Purpose
`fullstack-developer`	Shared	Implements a bounded phase across frontend, backend, or infrastructure under clear ownership.
`ui-ux-designer`	Engineer-specific	Designs or audits interfaces, responsiveness, accessibility, and design systems.
`code-simplifier`	Engineer-specific	Refines recently changed code for clarity and consistency while preserving behavior.
`docs-manager`	Shared	Audits and maintains technical documentation and development standards.

Verify, deliver, and learn

Agent	Origin	Purpose
`debugger`	Shared	Investigates failures, logs, behavior, data, or performance and returns a diagnosis.
`tester`	Shared	Runs and evaluates relevant tests, builds, coverage, and error paths.
`code-reviewer`	Engineer-specific	Reviews correctness, security, performance, edge cases, and maintainability.
`git-manager`	Engineer-specific	Handles scoped staging, commits, pushes, and Git handoff when explicitly requested.
`journal-writer`	Shared	Records significant failures, incidents, blocked dependencies, or architectural setbacks.

Discover and orchestrate Agents

Use ak agents list --kit engineer to inspect the resolved roster. After installation, the selected runtime projects and discovers Agents in its own way:

Begin with an outcome. Choose a Skill or describe the task you want to complete instead of treating the roster as a required sequence.
Let the workflow delegate bounded work. A Skill or your coding assistant can select a specialist when its role fits the current step.
Review the consolidated result. Delegation does not approve changes or external actions, and a person remains responsible for the final decision.

Claude Code uses its native subagent surface. Codex receives translated Agent definitions and AgentKit runtime integration. Cursor receives a native Agent projection. These are different orchestration surfaces, so a name appearing in the roster does not establish one direct invocation syntax across all three.

Runtime and model boundaries

Agent names and the purposes above describe the shipped roster, not guaranteed runtime parity. Model tiers can be mapped differently for each target, and the available provider or model can depend on the user's runtime setup. Treat an Agent's purpose as its working role, not as a promise of a particular model, identical output, or identical orchestration behavior.

For Codex in this release, the shared researcher Agent emits model = "gpt-5.4" and model_reasoning_effort = "medium" when its model resolves. If no model resolves, AgentKit omits both settings. This does not change the Agent roster, Claude Code behavior, or Cursor behavior.

For broader delivery differences, read Runtime adapters. If an installed Agent is not discoverable, follow Runtime cannot find a Skill or Agent before changing installation files.

Continue exploring

Return to the Engineer Kit overview for the complete package boundary and first workflow.
Read Kits, Skills, Agents, and Hooks to see how these components work together.
Use Installing kits to choose runtime, scope, delivery mode, and Skill selection.

On this page