A preconfig defines how an agent behaves: system prompt, tools, model. But it's stateless. Every session starts from scratch.

Promoting a preconfig to an agent gives it a persistent home: a private workspace, personal memory files, and a skills directory that follow it everywhere. The agent remembers your preferences across projects. It accumulates knowledge. It gets smarter the more you use it.

No magic out of the box

Jean2 doesn't ship opinionated agents. There's no pre-built coding assistant, no "expert researcher" persona, no plug-and-play personality that just works. The default preconfig is intentionally bare.

What you get out of Jean2 is proportional to what you put in. A vague system prompt and no capabilities enabled gives you a basic chat experience. A detailed system prompt, the right tools, capabilities turned on, and a promoted agent with persistent memory gives you something that feels like magic.

The platform provides the mechanics: memory that persists, skills that travel, sessions that accumulate knowledge, scheduling that runs autonomously. But the agent won't use them well unless you guide it.

You teach it how to learn. Tell it in your system prompt what's worth remembering and what isn't. Tell it when to create a skill, when to search past sessions, when to delegate to a subagent. The capabilities are tools, not instincts. The more you steer the agent toward using them, the more it pays off.

You provide the intent. The more you push it, the better it gets.

Promote a preconfig

Any preconfig can be promoted. The server creates a home directory and a hidden virtual workspace for the agent, then pre-enables every capability at low risk. No manual setup.

~/.jean2/agents/{agentId}/
├── home/                  # private workspace (hidden from normal lists)
│   └── .jean2/
├── skills/                # personal skills (travel across projects)
├── USER.md                # cross-project user preferences
└── MEMORY.md              # cross-project knowledge

Once promoted, select this preconfig in any workspace and the session is tagged with the agent's identity. The home directory is injected as an additional path, giving the agent filesystem access to its own persistent space. Demote the agent and the home is removed. The preconfig stays.

Two layers of memory

When a promoted agent runs in a workspace, it gets two memory tools instead of one:

Tool	Scope	Permission	Path
`memory`	This workspace	From workspace settings	`{workspace}/.jean2/`
`agent_memory`	All workspaces	No prompts, always allowed	`~/.jean2/agents/{id}/`

Workspace memory is for project-specific facts: repo conventions, build commands, architecture decisions. Agent memory is for cross-project knowledge: patterns you've refined, pitfalls you've hit, techniques that work everywhere.

Both are injected into the system prompt automatically. The agent knows the difference and saves accordingly.

Two layers of skills

Same pattern. A promoted agent sees skills from both locations, merged together:

Tool	Scope	Permission	Path
`skill_manage`	This workspace	From workspace settings	`{workspace}/.agents/skills/`
`agent_skill_manage`	All workspaces	No prompts, always allowed	`~/.jean2/agents/{id}/skills/`

If a workspace skill and an agent skill share the same name, the workspace one wins. Workspace skills are procedures for this project. Agent skills are personal workflows the agent has refined across everything it's worked on.

Session search across projects

Session search normally scopes to the current workspace. A promoted agent can additionally search by scope: "agent", returning sessions it has run in any workspace, past or present. The agent can recall work it did weeks ago in a completely different project.

The home workspace

Every promoted agent gets a virtual workspace at {agentId}-home. It comes with Memory, Skills, Session Search, and Scheduling all pre-enabled at low risk. It's hidden from the normal workspace list. The agent uses it as its personal sandbox, and it's available as an additional path in every session it runs.

What's next

Preconfigs: the personality bundles that agents are built from
Capabilities: memory, skills, session search, workflow, scheduling
Scheduling: let agents run tasks on a schedule