Workflows

Six recipes, start to finish. Each one is the shortest honest path through a real job — follow along once and the rest of the docs become reference.

Build a feature with two agents in parallel

  1. Code → Repositories → + Add Repository, pick your project.
  2. Agents → + New Agent: Claude Code, your repo, isolated worktree on. Prompt it with the first half of the feature (say, the backend).
  3. Launch a second agent the same way for the other half (the UI). Each has its own branch and folder — they cannot collide.
  4. Work on something else. When a session stops to ask, it badges Agents → Approvals — answer from there without switching tabs.
  5. As each finishes: Code → Review inbox → read the diff, Run tests, tick your UAT items, then Queue merge. The queue rebases, re-tests, and merges one branch at a time — a red step stops everything.
Save either session as a preset (hover its row → ☆) and next time this whole setup is one click.

Fix a GitHub issue, end to end

  1. Agents → + New Agent → in Start from issue paste #123 (or a Jira key like NBC-529) and hit Fetch. Title, branch name, and the opening prompt fill themselves from the issue and its comments.
  2. Launch. The agent starts already knowing the ticket.
  3. When it's done: Review inbox → tests + gate as usual → Open PR. The PR body carries the commits, stats and the verification record — and closes the issue automatically (Closes #123).

Run your agents on your own server

For when your project, files and logins already live on a Linux box.

  1. Servers → + Add server: a name and user@host (your normal SSH key/agent does the auth — AIOAT stores no passwords).
  2. The probe shows what's there: OS, load, tmux, and which agent CLIs are installed. Missing one? ⬇ install runs the official installer over SSH in a visible tab.
  3. Set the command (claude), the remote project directory, and tick tmux — the session then survives disconnects and reattaches next time.
  4. Keep share context ticked and the remote agent starts with your local workspace's shared context as its opening prompt — sent in-memory, never written to the server.
  5. Remote dev server running? ⇄ Forward port… (e.g. 5173:5173) and open http://localhost:5173 in the Preview pane.

Move an existing Claude Code setup into AIOAT

  1. Copy your .claude folder (subagents, skills, commands) and CLAUDE.md into the repo — or into %USERPROFILE%\.claude\ for machine-wide use. Every Claude session AIOAT launches picks them up as-is; nothing needs recreating.
  2. Knowledge → Importer → scan the repo → Import selected. Now the instructions live in AIOAT's context store too: they reach every agent's opening digest (Codex and Gemini included), show in the context graph, and drift is flagged if the files change later.
  3. Optional: Export canon as… writes the same guidance back out as AGENTS.md, GEMINI.md, or .cursorrules — maintain it once, ship it everywhere.

Delegate a whole goal to the Foreman

  1. Orchestrate → Foreman: describe the goal ("add rate limiting to the API and surface limits in settings"), pick the repo, set max workers.
  2. The Foreman reads the shared context and proposes a task plan. Nothing runs until you approve it — send it back with feedback if the split is wrong.
  3. On approval it spawns workers, each in its own worktree, visible as normal sessions. Watch progress in the Foreman pane or on Orchestrate → Canvas.
  4. Worker branches land in the Review inbox like any other — same tests, same gate, same queue.

A safe unattended run

Let agents work while you're away — with tripwires.

  1. Settings → Automation, in the workspace: agent done → run tests, and tests failed → send prompt "the tests are red — read the failures and fix them".
  2. Check Settings → General: notifications on, stall detection at a threshold you like — a silent agent flags itself instead of wasting the night.
  3. Checkpoints happen automatically at meaningful moments, so anything an agent breaks is one Code → Checkpoints rollback away (undoable — a safety checkpoint is taken first).
  4. Close the window if you want — sessions live in the detached host. In the morning: Home shows what ran, what it cost, and the run history; the Review inbox holds whatever earned a look.
Merges never happen unattended: the gate needs test artifacts on disk and your ticked checklist — and optionally a named human sign-off (workspace setting).

Not sure which model/agent? A/B it

  1. Palette → Agents: A/B test. Same task, two configurations — e.g. Claude Sonnet vs Opus, or Claude vs Codex, optionally with different prompt guidance per variant.
  2. Both run on twin branches. Compare the diffs and test results in the Review inbox; merge the winner, discard the loser.
  3. For a second opinion on any branch: Peer review hands the diff to the other installed CLIs as read-only reviewers.
← Core conceptsWorking with agents →