Improving Agents

Every agent you own has an Improve tab: the control center for changing what it does. You describe the change in plain English; Agentum prepares the work, protects your current version, and hands the actual coding to your chosen assistant — Claude Code or Codex — with everything it needs already loaded. When the assistant is done, you review the diff, accept it, and reload.

(Marketplace agents don’t have an Improve tab — they’re managed and sealed read-only. This page is about agents you created.)

The three stages

The Improve tab is a short wizard: Describe → Prepare → Hand off.

1. Describe

Pick what to change:

  • Idea cards — Agentum studies your agent and suggests improvements worth making (“AI ideas”); newly created agents start with curated quick-starts.
  • Fix banner — if the agent has been throwing errors, a banner offers to turn the actual error trail into a fix task. This is the fastest way from “it’s broken” to “it’s fixed.”
  • Tell me what to change — free text: “Make the digest include tomorrow, not just today, and send it 30 minutes earlier.”
  • Just open a session — power users can open the coding assistant on the agent’s folder with no prepared task at all.

2. Prepare

Agentum plays the change back to you before any code is touched: a short “here’s what I understood” review with a couple of clarifying questions if your request was ambiguous, then a plain-English “Here’s what will change” recap (with an optional visual preview of the intended result). In a hurry? Continue without review skips straight ahead.

Behind the scenes, Prepare also does the safety work:

  • A restore point of the agent’s current state is snapshotted — whatever happens next, you can get back here in one click.
  • The authoring kit (the full “how Agentum agents work” documentation and type definitions) is placed alongside the agent’s source, so the coding assistant knows the platform cold instead of guessing.

3. Hand off

Choose where the coding session runs:

Target What happens Best for
Inline An embedded terminal opens inside Agentum with your assistant already running in the agent’s folder, task loaded. The session is portable — full-screen, docked in the Improve tab, or riding the Sidekick rail — without restarting. Most people, most of the time. Also the only target that can offer to install the CLI if it’s missing.
Terminal macOS Terminal (or iTerm, if installed) opens with the same session. Toggles: skip the CLI’s own permission prompts; auto-close the window on exit. People with a beloved terminal setup.
Desktop app The task opens in the Claude desktop app via deep link. Claude Code only — hidden when Codex is selected. When you prefer Claude’s desktop UI.

There’s also Use another AI: copies a complete, self-contained prompt (task + context + platform rules) to your clipboard for ChatGPT, Cursor, or any other tool.

Which assistant launches — Claude Code or Codex — follows your Coding assistant setting; see Choosing Your LLM & Coding Harness.

After the assistant finishes

Changes made by the assistant land in the agent’s real source folder, but Agentum doesn’t blindly run them:

  1. Come back to the Improve tab — the Working copy panel wakes up and shows a read-only diff of everything that changed since your restore point, validated and classified by what it takes to apply (some changes are live instantly; others need an agent reload).
  2. Accept the changes (snapshotting a new version into History) — or Discard to throw them away and return to the restore point.
  3. Click Reload agent so the running system picks up the new code. Done.

Two more panels live alongside the Working copy, under Workshop:

  • Quick tweak — an in-app chat for one-liner edits (“change the send time to 8:15”) that proposes, verifies, and applies a patch without any external tool. It locks while unreviewed external edits are pending, so two editors never fight over the same files.
  • History — the agent’s version timeline. Every accepted change and every hand-off snapshot is a rollback target. One click restores any prior version.

Sidekick: ask about the agent

Next to the Improve machinery is a gentler companion: the Sidekick rail (the panel toggle in the agent’s toolbar). It’s a chat that answers questions about this agent — what it does, how to configure it, how to get more out of it — grounded in the agent’s guide, manifest, settings, and even its live data. Starter prompts get you going: “How does this agent work?”, “Walk me through setup.”

Sidekick is for understanding; Improve is for changing. The rail can also host your inline coding session, so both live comfortably side by side.

Practical tips

  • Small asks work best. “Add a snooze button” beats a five-part wishlist — you can run Improve as many times as you like, and each accepted change is its own rollback point.
  • Let the fix banner do the work when something’s erroring; it packages the real stack traces into the task.
  • Skip-permissions is a convenience with teeth. It lets the CLI edit without prompting you for each file — fine inside your agents folder, but it’s off by default for a reason.
  • Nothing is precious. Between restore points, the diff review, and History, the worst case is one click back to a working agent.

Next steps