Creating Agents
Creating an agent in Agentum is a guided conversation, not a coding project. You describe the outcome, answer a handful of quick questions, confirm a plan, and Agentum generates a working first version — then offers to hand it to a real coding assistant for polish. No files to create, no boilerplate to learn.
Click Create Agent in the sidebar to start. Your progress is saved continuously — you can leave mid-flow, use the rest of the app, even restart Agentum, and pick up where you left off.
Step 1: Describe it
The first screen asks “What do you want to build?” and gives you three ways in:
- I know what I want — type it in plain English. “Every weekday at 7am, send me a Telegram message with my calendar and the weather.” Starter chips below the box seed common shapes if you want a nudge. ⌘-Enter to start.
- Help me figure it out — a short three-question quiz about your work and pain points that synthesizes an agent idea for you, then feeds it into the same flow.
- Browse templates — jump to the Marketplace instead; sometimes the agent you want already exists.
Step 2: Shape it
Agentum asks a quick sweep of concrete questions — the when, where, and what-exactly that turn an idea into a build plan. Answer with taps and short text; you can re-answer anything without starting over. There’s no chat transcript to manage here — it’s a focused questionnaire, usually under a minute.
Step 3: Confirm the plan
You get a plain-English plan: what the agent will do, what integrations it uses (Telegram, Google Calendar, AI…), and how it’s put together. Edit answers loops you back; Looks good locks it in. You can attach a final note (“keep the messages short”) that rides along into the build.
Step 4: Build
Agentum recommends the path that works best in practice: draft it, then perfect it — the app generates a complete working draft itself, and then offers a hand-off to your coding assistant for refinements.
The build runs through generation, an automated code review, and a behavior check — you’ll see sample outputs of what the agent would actually produce (a preview of the morning message, for example) before anything is committed. When it passes, the agent’s files are written to your agents folder along with a GUIDE.md capturing what it is and how it works.
Step 5: Launch
The new agent is hot-loaded into the running system immediately — no restart, it appears in your sidebar and starts doing its job. The launch card offers:
- Open agent — see its panel.
- Improve this agent — jump straight into the Improve flow.
- Create another — back to step 1.
If you chose the draft-then-refine path, you land in the hand-off instead: your installed coding harness (Claude Code or Codex — see Choosing Your LLM & Coding Harness) opens on the fresh draft, ready for “make the message funnier”-level iteration with full context.
Where your agent lives
Your agents are plain folders of JavaScript under your agents folder — ~/Documents/Agentum/agents by default (changeable in System → Services → Agents folder). Each agent has a manifest.json, an entry file, and its GUIDE.md. You can read everything the builder wrote; nothing is hidden. Hand-authoring against the same format is fully supported — see Authoring Agents.
Two kinds of agents show up in your sidebar, and it’s worth knowing the difference:
- Your agents (badge: Editable) — anything you create here or by hand. Fully yours: Improve, Share/Export, direct file edits.
- Marketplace agents (badge: Marketplace) — installed from the catalog, managed and updated by Agentum, and sealed read-only (no Improve tab, no export). That’s deliberate: their updates would overwrite local edits. If you want a marketplace agent’s behavior with your own twist, create your own agent that does what you want.
If a build fails
Generation is checked at every stage — schema validation, cross-file consistency, a smoke test. If something can’t be repaired automatically, the flow tells you what failed and nothing broken is committed. Adjust the plan (usually the fix is making a fuzzy requirement concrete) and build again.
Next steps
- Improving Agents — iterate with Claude Code / Codex, review diffs, roll back
- Authoring Agents — the underlying format, for hand-coders
- Publishing to the Marketplace — share what you built