Agents
Custom AI agents that actually do work.
Skills, scoped permissions, persistent memory: any model you pick.
You decide what they can touch.
An agent only sees what you let it see. Pick folder by folder, computer by computer. Change your mind whenever.
Let it read your research folder. Let it edit your drafts folder. Keep your private notes off limits.
Take away a permission and the agent simply can't do that anymore. You can still go back and see what it did before.
Teach it skills.
Attach reusable skills: instruction bundles the agent reads when the task calls for them. Install one from the Hub or write your own.
Send it off with helpers.
A bigger agent can split work into pieces and hand them to smaller helpers, all running at the same time, all in one view.
It remembers what you taught it.
Pin a note. Drop in a file. The next time you ask, it picks up where you left off, across chats, across days.
- A real memory folder: things it reads first, every time you ask.
- Pinned notes for the corrections you keep having to make.
- Shared with the team so teammates benefit from how you've trained it.
- Versioned: roll back a bad change like any other file.
Custom GPTs are stuck with OpenAI's models and can't really touch your files or computers. Agents here can use any of the 200+ models, read the folders you allow, run things on a machine you own, and bring along smaller helpers, all in one view you can watch.
What you get
Mix GPT, Claude, Gemini, Grok, Llama, Qwen, DeepSeek. Swap the model without rebuilding the agent.
Drop in web search, code execution, files, computers: only the capabilities you opt in to.
Grant per-folder file access, per-computer command access, per-credential identity. Revoke anytime.
Agents remember prior runs across conversations. Pin notes, attach files, share between team members.
An agent can spawn focused subagents for parallel research, drafting, or code review. You see every step.
Keep an agent for yourself, share with your workspace, or publish it to the Hub for the world to install.
Who it's for
An agent with web search + a paper folder writes literature reviews you can cite.
A code-review agent reads your repo, runs the test suite on a connected computer, posts findings to your tracker.
A weekly-report agent pulls metrics, drafts the summary, drops it in the shared workspace folder.
An editing agent learns your voice from past drafts and applies it consistently across new work.
How it works
- 1Describe the agent
Name it, write a system prompt, pick its default model.
- 2Grant skills
Tick which tools it can use: web, files, computers, code, subagents.
- 3Scope its access
Pick the folders it can read, the computers it can command, the credentials it can mint tokens against.
- 4Hand it work
Mention it from any chat, schedule recurring runs, or invoke it from the CLI.