How Skills Work Differently on the General Agent vs. Custom Agents
Last updated: March 21, 2026
The general agent (your default chat at gumloop.com/chat) and custom agents handle skills in fundamentally different ways — different discovery methods, different attachment rules, and different behavior when skills are created during a conversation. Understanding these differences helps you set up skills correctly for each agent type.
Side-by-Side Comparison
Behavior | General Agent | Custom Agent |
Which skills can it see? | All skills in your workspace — automatically | Only skills explicitly attached to it |
How does it find the right skill? | Semantic search — matches your request against skill descriptions dynamically | Scans the list of attached skill names and descriptions in its system prompt |
Are skills listed in the system prompt? | No — that wouldn't scale for large workspaces. Uses a | Yes — each attached skill's name and description is included (~50–100 tokens per skill) |
Do I need to attach skills manually? | No — every skill in your workspace is discoverable | Yes — go to the agent's config → Skills → + Skill |
What happens when a skill is created during a conversation? | Added to your workspace library. Not auto-attached to any agent. | Auto-attached to this specific agent immediately. |
Scales to many skills? | Yes — semantic search works regardless of how many skills you have | Practical limit — each skill adds tokens to the system prompt, which can crowd the context window |
How the General Agent Finds Skills
The general agent uses a dedicated skill_discovery tool. When you send a message, the agent can query this tool with a description of what it needs, and the tool returns matching skills using semantic search (embedding-based matching against skill descriptions).
This means:
You don't need to attach skills — they're all searchable
The agent can find skills even if your wording doesn't exactly match the skill name
The agent can also filter skills by integration (e.g., "find skills related to Google Sheets")
There's no token overhead from having many skills, since they're not listed in the system prompt
How Custom Agents Find Skills
Custom agents take a different approach: every attached skill's name and description is injected directly into the system prompt. When you send a message, the agent reads through this list and decides which skill (if any) to load.
This means:
You must attach skills manually in the agent's configuration
The agent only knows about attached skills — it cannot discover skills that exist in your workspace but aren't attached
Each attached skill costs ~50–100 tokens in the system prompt, so attaching dozens of skills adds up
Skill descriptions need to be specific enough for the agent to distinguish between them
When a Skill Is Created During a Conversation
On a custom agent
If the agent creates a new skill during a conversation (e.g., you say "remember this process as a skill"), the skill is automatically attached to that agent. It will be available in future conversations with that same agent without any manual setup.
On the general agent
The skill is saved to your workspace library, but it is not auto-attached to any specific agent. It becomes discoverable by the general agent through semantic search in future conversations, but if you want a custom agent to use it, you'll need to attach it manually.
Which Should I Use?
Use Case | Best Choice |
You have a large library of skills and want flexible discovery | General agent |
You want tight control over exactly which skills are available | Custom agent |
You're building a focused agent for one job (e.g., "Sales Assistant") | Custom agent with 3–10 relevant skills |
You want skills to auto-improve through conversation corrections | Either — both support skill editing when the toggle is enabled |
You have many skills that overlap in description | Custom agent — you can control which ones are visible and use the system prompt to disambiguate |
Related Docs
Still Need Help?
If this didn't resolve your issue, reach out to support at support@gumloop.com.