Docs

Agents

Edit an agent

Tune how an agent sounds and works, and correct what it has learned.

Edit an agent when you keep correcting the same habit, when the project changes under it (a new eval set, a moved data folder), or when its tone doesn't fit your team. It matters because agents act on what they've written down: a stale memory note or an out-of-date system prompt is followed faithfully, into the wrong folder.

Every agent has a profile: a card that says who it is, and four tabs with what it keeps. Open it from the Agents page by clicking the agent's card, or hover the agent in the sidebar's Agents section and click the profile button at the end of its row (Open Kestrel-17 profile, for example). Clicking the agent's name itself opens a chat with it.

Kestrel-17's profile card with the personality editor open, Reset to built-in, Cancel and Save below it, and the AGENTS.md system prompt row
Personality and system prompt sit on the profile card. The tabs below hold what the agent has learned.

What's on the profile

PartShows
CardName, @handle, a Chat button, and counts for each tab
AboutThe agent's role, a role badge, and Built-in or Workspace (an agent you created)
PersonalityHow it sounds in chats, channels and comments
System promptIts AGENTS.md: identity, behavior and working instructions
MemoryNotes about this workspace: layout, conventions, decisions, past fixes
SkillsHow-tos it wrote for itself, like running your eval suite
ExtensionsCode it added to extend its own tools
HeartbeatChecks that wake it on a schedule or when something happens

Click a count on the card to jump to that tab. Refresh in the card's corner reloads the profile after the agent has been working.

Change its personality

Personality is a few short phrases that set the agent's voice. Change it when an agent runs too chatty, too terse, or off-tone for your team.

  1. Click Edit personality

    The pencil next to the personality text on the profile card.

  2. Edit and click Save

    Up to 600 characters. Two or three phrases work best: “blunt and exact; asks for seeds and error bars; short replies”. The card then reads Customized for this workspace.

Once you have changed it, the editor also shows Reset to built-in, which returns to the personality the agent started with.

Built-in agents keep one personality in every workspace. Their card reads Managed by arkenOS and the same in every workspace, with no pencil. Personality editing is for agents you created.

Change its system prompt

The system prompt is the agent's standing instructions, read at the start of every task. Click AGENTS.md under System prompt to open it as a document, then edit it like any page. It saves a moment after you stop typing, and ⌘S saves right away.

This is the place to fix a habit you keep correcting:

Instead of

Be careful with evals.

Try

Report every eval as the mean of three seeds, with the worst seed beside it. Never compare runs that used different held-out files.

The agent edits AGENTS.md too as it learns, so read what's there before rewriting it. Agents without one show Runs on its built-in instructions.

Correct what it has learned

Click any file in Memory, Skills, Extensions or Heartbeat to open it. Each tab has its own page: Memory, Skills, Extensions and Heartbeat. Edits save the same way, and the agent uses the corrected version from its next task.

  • Skim Memory after a big change. A moved data folder or a new eval set makes old notes wrong. Fix the note, or delete the line.
  • Check skills it relies on. If an eval skill still points at last month's held-out file, every run built on it is off.
  • Prefer telling it in chat for small things. “Remember that the held-out set moved to evals/v2/” updates memory without opening a file.
  • Memory belongs to the workspace. Each project teaches its agents separately, so a fix here stays here.