Agent Capabilities
Memory, skill management, session search, and workflow orchestration. The features that make your agent smart over time.
Location: Workspace Settings → Capabilities tab group
Capabilities are what turn your agent from a chatbot into something that gets smarter the more you use it. Memory lets it remember. Skills let it learn new behaviors. Session search lets it recall past work. Workflow lets it delegate complex tasks to parallel subagents.
All four are off by default. Toggle them on per workspace. Each capability (except Workflow) has a risk level that controls when the agent asks before writing.
Memory
Gives the agent a persistent memory store. Facts, decisions, architecture notes, preferences. Anything worth remembering across sessions gets saved and loaded into context automatically.
Without memory, every session starts from zero. The agent re-learns your codebase, re-discovers your conventions, re-asks the same questions. With memory, the third session is smarter than the first. The tenth is smarter than the fifth.
Example: You tell the agent "we use pnpm, not npm" in session one. Two weeks later, in a new session, it already knows. No AGENTS.md needed. The agent decided it was worth remembering and saved it.
Enable: Workspace Settings → Capabilities → Memory. Toggle on, pick a risk level.
Skills
Lets the agent manage its own skills. Skills are SKILL.md instruction sets. When enabled, the agent can read, create, update, and patch them using the skill_manage tool. No manual file editing required.
Skills are how the agent learns specialized workflows. Code review checklists, deployment procedures, debugging playbooks. Instead of you writing every skill by hand, the agent creates them as it discovers patterns in your work. The agent essentially programs itself.
Example: The agent notices you keep asking it to follow a specific commit message format. It creates a skill for it. Next session, the skill is there. You never wrote a line.
Enable: Workspace Settings → Capabilities → Skills. Toggle on, pick a risk level.
See Skills for the SKILL.md file format and how skills are loaded.
Session Search
Lets the agent search through past sessions in the current workspace. It can recall what was discussed, what was decided, and what tools were run in previous conversations.
Sessions accumulate. After a week of working in a workspace, you might have 20 sessions. The agent can't remember any of them unless you enable search. With it, the agent can look back: "what did we decide about the auth module last Tuesday?"
Options:
- Risk level: approval threshold for search operations
- Include tool results: whether search returns include past tool outputs. More context, more tokens.
Enable: Workspace Settings → Capabilities → Session Search. Toggle on, pick a risk level, optionally enable tool results.
Workflow
Enables the workflow orchestration tool. The agent can decompose a complex task into parallel subtasks, fan them out to multiple subagents simultaneously, and synthesize the results.
Some tasks are inherently parallel. "Review these 5 files for bugs" runs 5x faster if each file goes to a separate subagent. Workflow makes this automatic. The agent decides what can be parallelized, spawns the subagents, and combines their work.
Example: You ask the agent to audit a codebase for security issues. Instead of reading every file sequentially, it spawns 4 subagents, each scanning a different directory, then compiles a single report. What would take minutes takes seconds.
Enable: Workspace Settings → Capabilities → Workflow. Simple toggle. No risk level needed since workflow itself doesn't write to disk.
See Subagents for how the orchestration works under the hood.
Risk levels
All capabilities that involve writing (Memory, Skills, Session Search) share the same risk level options. Higher risk means more friction, more control. Start with Medium and adjust based on your comfort level.
| Level | Behavior |
|---|---|
| None | Auto-approve all writes. The agent never asks |
| Low | Only ask for bulk changes |
| Medium | Ask for each write |
| High | Require explicit approval |
| Critical | Always confirm with full details |
What's next
- Skills: how SKILL.md files work
- Subagents: how the orchestration works
- Workspace Settings: all workspace config