Talking to the agent

Once the agent is on for your workspace, you can talk to it from any page. Press ⌘J (Ctrl+J elsewhere), click Ask the agent in the top bar, or click Assistant in the sidebar. In a terminal, run loomctl agent. You need agent:use, which every member has, and the agent works with your permissions, not its own. If it isn't on yet, start with the built-in agent.

Ask it things like "why did the last deployment of api fail?" or "what's the latest commit and release of shop?". It reads live state with its tools, links what it mentions, and asks before anything risky.

The panelCopy Anchor Link

The agent lives in a panel beside your work, so you never leave the page you're on. ⌘J opens it and closes it again, and the command menu has it as Ask the agent. The panel can sit in three places, picked from the layout button in its header:

On a phone the panel covers the screen. Whatever you pick, Loomplane remembers it in your browser, along with the width, whether the panel was open, and the conversation you had open, so a reload puts you back where you were. If you close the panel while the agent is still answering, the answer keeps coming in, and it's there when you open the panel again.

Old links to /assistant still work. They open the panel at full width, on the conversation the link names.

The orb in the panel's header changes with what the agent is doing. It turns slowly when the agent is idle, faster while you type and faster again while the agent works out an answer. A ring goes round it while a tool runs, it pulses while the answer streams in, it turns amber when the agent is waiting on you, and it goes red and stops when a turn failed.

What a message is aboutCopy Anchor Link

Above the box, two chips say what you're asking about.

The first is the conversation's scope: a project, or the whole workspace. A new conversation starts on the project of the page you're on, and you can pick another one, or the whole workspace, before you send. After the first message the scope is fixed, and the chip shows it. To change it, start a new conversation.

The second is the page you're looking at, such as "Service web in api, production" or "Deployment plan_42". It goes with your message, so "why is this slow?" on a service's page is about that service, and the agent reads that service before it answers. Click the × to leave the page out of a message; it stays out until you move to another page, and the dashed chip puts it back. The page only goes along when it fits the scope: a page from another project doesn't. Each message you sent shows the page it came from, as a small link above it.

The suggestions on a new conversation follow the page too. On a service you get questions about that service, on a deployment questions about that deployment.

Enter sends and Shift+Enter starts a new line. Up in an empty box brings back the last message you sent. One message is answered at a time: while the agent is on one, the box is off and its button is Stop, which cancels the task. If a message goes in anyway (from another tab, say), the server says no with "The agent is still on your last message. Wait for it to finish, answer what it asked, or stop it." The same message comes back from the API as a 409.

Your conversationsCopy Anchor Link

The clock button in the panel's header opens your conversations. You can search them by title or project, filter by Needs you, Working or Failed and by project, and group them by project or by date (the button beside the search switches between the two). Each group folds away.

Hover a conversation to pin it, or open its menu to rename or delete it. Pinned conversations sit at the top, in every browser you sign in from. Deleting one takes it out of your list for good, and stops the agent if it's still working on it. Only the person who started a conversation can pin, rename or delete it.

What answers look likeCopy Anchor Link

The agent writes Markdown, and the panel renders a few things beyond plain text:

• Code comes in blocks with the language named, colours for the common ones (shell, YAML, JSON, Go, HCL, Dockerfile, SQL and a few more) and a copy button.
• Tables put a status word such as Failed or Live beside a coloured dot, line numbers up on the right, and copy as CSV.
• Links to Loomplane pages stay in the workspace and show what kind of page they open. Links to commits, releases and repositories open in a new tab.
• Figures can come as headline tiles, as a chart (bar, horizontal bar, stacked, line or area), or as a timeline of releases. Every chart opens as a table too, and copies as CSV.
• A note, tip or warning comes as a coloured callout.

The agent only charts numbers its tools gave it. If a chart it wrote can't be drawn, the panel shows what it wrote instead, and says so.

Steps and how risky they areCopy Anchor Link

Each tool the agent uses shows as one quiet line between the words, such as "Listed deployments" or "Checked the repositories". Steps taken back to back fold into one line that says what they did and how many there were, and it stays open while the agent is on it. Click a line to see the tool, its arguments and what it returned.

A step that's more than a read carries a label for how risky it is. There are six, one for each class on What needs a person.

A step that failed says so, a step that was refused says "refused", and a step that never got a result says "no result". A turn you stopped says "Stopped."

Commits and releasesCopy Anchor Link

Ask the agent what a repository is on and it reads the repository's host: the head of the branch the project tracks (or the default branch), the latest release (or the newest tag, on Bitbucket, which has no releases), and the commit each environment runs, with links to all three. It works for GitHub, GitLab (self-managed too) and Bitbucket, through the same connections your deployments use. A repository the host won't show is reported as such, and the others still answer.

When the agent asks you somethingCopy Anchor Link

If the agent needs an answer, a question appears in the conversation with its options, and you answer it there. The same request shows on Deployments, Agent with everything else the agent is waiting on, and answering in either place settles both. Plan approvals, secrets and connections still go through their own cards, with the same rules as before.

What a safety stop looks likeCopy Anchor Link

Loomplane scans every answer before it counts as an answer. If one holds a credential, such as an API key or a token, the agent stops that answer and the conversation shows "Loomplane's safety checks stopped this answer." Tool calls in that answer don't run. Other personal data, an email address for instance, is redacted in place and the answer goes on.

This assumes shield's stock output profile. If you've changed shield's profiles, the line between a stop and a redaction is yours.

Who can read and add to a conversationCopy Anchor Link

A conversation is private to the person who started it, with one exception.

Someone with ai:manage, which admins have, can read everyone's conversations in the workspace, for support and for audit. The API lists them with GET /api/v1/agent/conversations?everyone=true. Anyone else asking for a conversation that isn't theirs is told it doesn't exist.

Only the person who started a conversation can add to it, rename it, pin it or delete it. An admin can read it and stop its running task, and can't post in it. If you want to carry on from someone else's thread, start your own.

A message is up to 16,000 characters.

Report an answerCopy Anchor Link

If an answer is offensive, harmful or wrong, report it. Under every finished answer there's a flag button, Report answer. Pick a reason (Offensive, Harmful, Inaccurate or Something else), add a note of up to 1,000 characters if you want to, and send it. The panel says "Thanks. The report went to whoever runs this Loomplane."

Anyone who can read the conversation can report one of its answers: the person who started it, or someone with ai:manage. Report the same answer again and you update your report instead of adding a second one. Over the API, it's POST /api/v1/agent/conversations/{id}/turns/{task}/flag.

A report keeps the question, the answer (its first 8,000 characters), the reason and your note, plus the ids that say where they came from. If you have ai:manage and report an answer in someone else's conversation, their question goes with it. The workspace's audit trail records that a report was made, with those ids and the reason. It never holds the question, the answer or your note.

The report goes to the platform owner, not to your workspace's admins. On your own server that's whoever set it up, and on app.loomplane.com it's us. The owner reads reports under Platform, Flagged answers, filters them by Open, Dismissed or All, and dismisses one once it's dealt with. Other platform admins don't see the page. A dismissal lands in the audit trail of the workspace that reported, under the owner's name.

A report doesn't expire. It stays after the conversation is deleted, so the owner can still review it, until the owner removes it. To have one removed, ask the owner. On app.loomplane.com that's us, at the address under Your rights in the privacy policy.

From the terminalCopy Anchor Link

loomctl reads LOOMPLANE_URL and LOOMPLANE_API_KEY, as it does for every command. The agent acts as the person who made the key.

loomctl agent                       # a conversation at a prompt
loomctl agent --project shop        # about one project (slug, name or ID)
loomctl agent --conversation <id>   # pick one up again

At the you> prompt, type a message and press Enter. Text streams back, and each tool shows on stderr with its risk, such as · list_deployments (read). A few slash commands:

A question from the agent shows at an answer> prompt with numbered options. Type a number or your own words. An empty line leaves it pending, and you can answer later with loomctl agent requests answer <id> <text>. Ctrl-C stops the task.

loomctl agent conversations [--project <ref>] [--json] lists yours. loomctl agent conversations rename <id> <title> renames one and loomctl agent conversations delete <id> deletes it, stopping the agent if it's still on it.

One question, in a scriptCopy Anchor Link

loomctl agent ask sends one message, prints the answer on stdout and exits.

loomctl agent ask "what changed in the last deployment of shop?" --project shop

Anything that isn't the answer, the steps and the reasons it failed, goes to stderr, so you can capture the answer cleanly. --json prints every event as one JSON line instead. --conversation <id> adds to a conversation you started.

Without --no-wait, ask waits with the agent, so a question it puts blocks a pipeline until someone answers. In CI you almost always want the flag:

- name: Summarise the deployment
  env:
    LOOMPLANE_URL: ${{ vars.LOOMPLANE_URL }}
    LOOMPLANE_API_KEY: ${{ secrets.LOOMPLANE_API_KEY }}
  run: |
    status=0
    loomctl agent ask "why is the latest production deployment of shop slow to start?" \
      --project shop --no-wait > summary.md || status=$?
    if [ "$status" -eq 3 ]; then
      echo "The agent needs a person. See: loomctl agent requests" >&2
    fi
    exit 0

The final exit 0 is a choice: a summary is rarely worth failing a build for.