Describe an agent
Write the brief the way you would hand the job to a colleague — the goal, the boundaries, and what done looks like. An agent whose only tools are draft editors turns that brief into something you can read before anything runs.
The planning conversation
New Task Agent opens a planning session. You describe the job; the planning agent asks what it needs to know — which services, what output, what to do at the edges — and drafts an execution plan as you talk. The draft renders twice: as readable markdown in the chat, and structured in the Agent Blueprint panel beside it.
The planning agent’s editing tools are deliberately draft-only. It can propose and revise steps all day; it cannot save, schedule, or run anything. The only commit point is you.
What you edit where
You change the plan by replying — “drop step 3”, “post to #ops instead”, “only include deals over $10k”. You change the configuration directly in the panel:
| Panel section | What you set |
|---|---|
| Title | The agent’s name. |
| Execution Model | The model runs use — independent of the planning model — plus reasoning effort. |
| Execution Plan | Read-only view of the drafted steps. |
| MCP Servers | The services this agent can reach. |
| Schedule | Frequency, day, and time — or on-demand. |
If the plan needs a service you haven’t connected, the planning agent says so and can offer the connection inline — see Credentials.
Saving
Save Agent validates the plan (every dependency must exist, no cycles), checks that you are allowed to run everything in it — the agent, each provider, each tool — and saves it as a Task Agent. Run Now does the same and immediately queues a run, which is the right first move: watch one run end to end before trusting a schedule. If saving is blocked by policy, the error names exactly what you lack; see Troubleshoot a run.
Files you attach during planning become the agent’s base context — every run can read them. Attachments are read-only reference material: an agent can read a spreadsheet you attach, but cannot edit it or hand back a modified copy.
Convo Agents
New Convo Agent opens a different interview. Instead of steps and schedules, the designer asks about five things: purpose, voice and tone, decision rules, tools, and output style. There is no cron and no plan — a Convo Agent always runs as the person chatting with it, in the moment.
When you Save agent, Lunen synthesizes the agent’s system prompt from the entire design conversation — not just your last message — so everything you agreed on in the interview makes it into the agent. The saved agent lives under Convo Agents, where anyone with access can start a New chat.
A Convo Agent’s tools authenticate as whoever is chatting. Two members asking the same agent the same question can get different answers when their connected accounts see different data — that’s data scoping working as intended.
Drafts
A planning session you walk away from is not lost — it appears under Task Agents filtered as a draft, and reopens exactly where you left it. An agent that was already saved can be reopened for edits with Edit Plan on its detail page, which returns you to the same planning conversation.