Custom Agents: VS Code & CC double down, Cursor removes its own

Ideas Feature Requests
1

Feature request for product/service

Cursor IDE

Describe the request

TL;DR

- Cursor 2.1 removed Custom Modes, which were the basic building block for custom agents in Cursor.
- VS Code and Claude Code are both doubling down on configurable custom agents / subagents with file-based definitions, metadata, and org-level sharing.
- Without a custom agent story, Cursor risks falling behind in agent-first workflows. Please consider restoring and evolving Custom Modes into proper custom agents.

I’m a heavy Cursor user and really appreciate how far the product has pushed AI-assisted coding. That’s exactly why the removal of Custom Modes in Cursor 2.1 feels like a major regression – especially for people trying to build agent-first workflows, not just use “AI autocomplete”.

Agent-first workflows are where the ecosystem is heading

Across the ecosystem, more teams are experimenting with letting agents handle the full software lifecycle:
requirements → design → implementation → refactoring → code review → maintenance.

In that world, a single generic chat box is not enough. We need stable, reusable custom agents with:

  • Their own system prompts (persona, rules, constraints)
  • Carefully scoped tool access
  • Optional model selection per agent
  • The ability to be shared and versioned across a team / org

Cursor’s old Custom Modes were already a step in this direction (fixed tools + system prompt + model per mode). Removing them goes directly against this agent-first trend at the exact moment competitors are doubling down on custom agents.

VS Code is clearly betting on Custom Agents

VS Code didn’t just keep its “custom chat modes” – it renamed and upgraded them into first-class Custom Agents:

  • Custom agents are defined via .agent.md files with rich metadata like name, description, argument-hint, model, tools, handoffs, and target, so they can run consistently across VS Code, GitHub Copilot Cloud Agents, and CLI Agents.
  • Chat modes were formally renamed to custom agents and moved into .github/agents with .agent.md definitions, making them shareable and discoverable inside repos.
  • They are adding organization-level agents you can share across a GitHub org (experimental), so companies can standardize agent personas and workflows.
  • Custom agents integrate with background agents (git worktrees / isolated workspaces), so multi-agent orchestration is part of the core editor story.

All of this is backed by dedicated docs: VS Code Custom Agents documentation.

In other words, VS Code is explicitly treating custom, composable agents as a long-term pillar of its Copilot experience, not as a nice-to-have extra.

Claude Code’s subagents are another strong, configurable agent system

At the same time, Claude Code has Subagents, which are also fully configurable agents defined as Markdown with YAML frontmatter – conceptually very close to VS Code’s .agent.md approach:

  • Each subagent has a name, description, its own system prompt, optional tools list, model, permissionMode, and skills.
  • You can define project-level and user-level agents via .claude/agents/*.md and ~/.claude/agents/*.md, share them with your team, and keep them in version control.
  • There’s also CLI-based configuration (claude --agents '{ ... }') to define subagents dynamically for scripts or one-off workflows.
  • The docs explicitly position subagents as reusable, specialized personas for tasks like code review, debugging, data science, etc. See: Claude Code Subagents documentation.

Subagents, combined with skills and plugins, are one of Claude Code’s killer features. They make it natural to build multi-agent workflows that teams can standardize on.

If Cursor no longer has even a basic notion of custom agents (configurable system prompts + tools + model per persona), it’s hard to see how Cursor can stay competitive for agent-first users as VS Code + Copilot and Claude Code keep pushing forward.

What Cursor users lose without Custom Modes / custom agents

Without any built-in concept of custom agents or modes, a lot of very practical workflows get much clumsier:

  • Team-level personas

    • A “Security Reviewer” agent that always enforces your org’s security checklist.
    • An “Architecture Guardian” that pushes back when changes violate design principles.
      Today, each developer has to manually paste long instructions into chat instead of just selecting a stable, named agent.

  • Role-based agents in a pipeline

    • Planner → Implementer → Refactorer → Reviewer, each with different prompts and tools.
      With custom agents, this becomes a one-click handoff. Without them, it’s all manual prompting and copy-paste.

  • Repo-specific agents

    • An agent that “knows” your monorepo’s structure, internal frameworks, coding conventions.
      In VS Code and Claude Code, this lives as .agent.md / .claude/agents/*.md in the repo, so the whole team just gets it. In Cursor right now, there’s no equivalent.

Cursor’s previous Custom Modes at least gave us a minimal version of this: fixed tools + prompt + model. Removing that takes away a key building block for agent-first workflows.

Constructive suggestions for Cursor

I’m not asking Cursor to clone every VS Code / Claude feature overnight. But there’s a clear incremental path that would dramatically strengthen Cursor’s agent-first story:

Short term – restore a minimal Custom Mode / custom agent primitive

  • Bring back something equivalent to the old Custom Modes:
    • Per-mode system prompt, default tools, and model selection.
    • Usable at both user and workspace scope, with quick switching in the UI.
  • This alone would unblock many workflows that people already had built on top of Cursor.

Mid term – move toward file-based, shareable agents

  • Introduce a file-based agent definition format (for example .cursor/agents/*.md, or even compatibility with .agent.md):
    • Metadata like name, description, argument-hint, model, tools, handoffs.
  • Support workspace, profile, and ideally org-level agents, so teams can standardize agent personas across projects – similar to VS Code’s .github/agents and Claude Code’s .claude/agents.

Long term – make Cursor the leading agent-first IDE

  • Integrate custom agents deeply with:
    • Background / cloud / CLI agents.
    • Git operations, multi-file edit tracking, and review workflows.
  • Go beyond what VS Code / Claude do in UX and ergonomics. Cursor is already ahead in many “AI-native editor” aspects; strong custom agents would multiply that advantage.

From a competitive perspective, the entire ecosystem is moving toward stronger custom agents and multi-agent orchestration. Cursor is in a great position to lead here, but removing Custom Modes without a stronger replacement sends the opposite signal.

Why I’m posting this

This isn’t meant as a rant. It’s feedback from someone who genuinely wants to keep investing in Cursor as my main agent-first IDE, instead of being forced to move more of my workflows into VS Code or Claude Code just to get robust custom agents.

I’d really appreciate if the team could:

  • Share how you’re thinking about custom agents / Custom Modes on the roadmap, and
  • Consider restoring at least a minimal version, or introducing a new, more powerful custom agent system inspired by what’s clearly working in VS Code and Claude Code.

If Cursor wants to stay the go-to IDE for serious AI-driven development, custom agents are not optional – they’re foundational.

6 Likes
2

An excellent business analysis! Yet another reason to bring back and expand Custom Modes!!

3

You can just use commands which are 80% the same as custom agents. The only real difference is not being able to scope the tools strictly (other than saying not to use them and hoping the agent respects that).

I do agree though that custom agents would be the preferred path. I’d also like the ability to hide the built in agent modes because I never use them and prefer my own scoped commands. They recently added another agent mode which was pretty pointless for me since I already have a command that covered it.

It would be much better to expand the flexibility and customization of cursor than add prebuilt magic modes that are black boxes.

4 Likes
4

Custom Modes allowed locking down tools, locking down models, ensure overriding system messages were applied on every request instead of only when called out explicitly, mitigated wasted tokens due to unnecessary tool calls and processing, allowed the development of custom workflows with reliable, repeatable results… The list goes on. So for me.. commands does about 50% what was needed , but also gives me zero piece of mind the ide level lockdowns facilitated.

1 Like