# How an ADE works: from task to shipped code in Tallos > An ADE works by giving every task its own isolated workspace, running an AI coding agent in it, and routing the result through your review before anything merges. In Tallos, you connect the agents you already use, start each task in a git worktree, run several agents in parallel, preview and review their diffs, then open a PR or discard the copy. - Canonical: https://runtallos.com/how-it-works - Português: https://runtallos.com/pt/como-funciona.md - Author: [Roberto Rocha](https://www.instagram.com/robertorochamkt/) - Updated: 2026-09-30 ## Key facts - One task = one git worktree, one branch, its own terminals - Agents start in manual mode and ask before running commands - Diff notes go back to the agent in one batch - Squads: a leader agent plus up to 12 members - Nothing merges until you commit, open a PR and merge ## How does an ADE work? An **ADE (agentic development environment)** works as a loop: a task becomes an isolated workspace, an agent works inside it, and you review the diff before anything reaches your main branch. The agent writes the code; the ADE handles isolation, sessions, status, preview and review. Tallos is that loop as a desktop app for macOS 13+ and Windows 10+. It has no model of its own: it runs the agent CLIs you already use, such as Claude Code, Codex or Gemini CLI, on your existing subscriptions and API keys. New to the term? Start with [what an ADE is](https://runtallos.com/what-is-an-ade). ## What happens between a task and shipped code? Every piece of work in Tallos follows the same nine steps. You step in to describe, answer and decide. 1. **Connect your agents** — Open Settings → Connections. Each agent card installs the CLI with the vendor's verified command, opens its sign-in or saves an API key for that agent only, and lists its models. See [agent connections](https://runtallos.com/features/agent-connections). 2. **Add a project** — Choose a folder with a git repository, clone one from a URL, or create a new project. Tallos uses your default branch as the base every new workspace starts from. 3. **Create a task** — Click **New workspace**, pick the project and the agent, and name the task (or leave it blank). Optionally link a GitHub, GitLab, Linear or Jira item. The worktree and branch are created in the background. 4. **Describe the work** — Claude Code, Codex and Grok CLI open in a chat view; agents without chat support, such as Gemini CLI or OpenCode, open in the terminal. Say what to change and how you will check it. 5. **Run more in parallel** — Create another workspace for a different task, or give the same task to a second agent with **Try another agent**. Each one gets its own copy of the repository. 6. **Watch and answer** — Status icons show who is working, who needs you and who is done. Agents start in manual mode, so they ask before running commands. 7. **Preview the result** — Open your dev server in the workspace's built-in browser. Click an element on the page to send its markup, styles and a screenshot to the agent. 8. **Review the diff** — The diff compares the workspace with its start branch. Leave notes on lines, then **Send notes to** the agent in one batch. Repeat until it is right. 9. **Keep or discard** — To keep the work, commit, push and **Create PR** from Source Control, then merge. To drop it, delete the workspace: Tallos removes the folder and its branch. > See the loop on your own repo. Tallos runs on macOS and Windows with the agents you already use. → https://runtallos.com/signup ## Why does every task get its own git worktree? Every task gets its own git worktree so agents never overwrite each other. A **git worktree** is a second working folder of the same repository, with its own branch and files on disk. Two agents can change the same file in two workspaces, and nothing collides until you choose what to merge. A new worktree is a clean checkout, so gitignored files such as `.env` are missing; list the ones to copy in a `.worktreeinclude` file at the repo root. A workspace can start from the base branch, another branch, a commit or a remote branch. Details in [parallel workspaces](https://runtallos.com/features/parallel-workspaces). > **Isolation covers files, not permissions** — A worktree keeps tasks apart; it does not sandbox the agent from your computer. That is why Tallos starts agents in manual mode. Letting agents run commands without asking is opt-in, under Settings → Agents → Agent Permissions. ## How do you run and watch several agents at once? You run several agents at once by opening one workspace per task and arranging their panes side by side. Drag a tab to a pane's edge to split it, and mix agents, dev servers, tests, diffs and browser tabs in one layout. - **Chat or terminal.** Supported agents open in a readable chat view with messages, tool calls and question cards. **View terminal** shows the raw CLI; **Back to chat** returns. See [native chat](https://runtallos.com/features/native-chat). - **Status at a glance.** Each session shows working, needs you, done or idle in the sidebar, so you only open the one that needs an answer. - **Usage readouts.** The status bar shows usage and rate-limit windows for agents such as Claude Code, Codex and Gemini. - **A browser per workspace.** Each agent's app stays next to its code. How many agents make sense depends on your machine, your agent plans and how much you can review. See [how to run AI agents in parallel](https://runtallos.com/run-ai-agents-in-parallel). ## What happens where in Tallos? Each stage of the multi-agent coding workflow has one place in the app. | Stage | Where in Tallos | What you do | | --- | --- | --- | | Connect | Settings → Connections | Install agents, sign in or save an API key, see models | | Pick work | Tasks view (GitHub, GitLab, Linear, Jira) | Start a workspace from an issue, PR or ticket | | Isolate | New workspace: a git worktree and branch | Choose the project, start point and agent | | Instruct | Agent pane: chat view or terminal | Describe the work, answer questions | | Run | Terminal splits | Dev servers, tests and more agents side by side | | Monitor | Sidebar status, notifications, status bar | Answer whatever needs you | | Preview | Built-in browser, one per workspace | Click an element and send it to the agent | | Review | Diff viewer | Leave line notes and send them back in one batch | | Ship | Source Control and the PR tab | Commit, push, open the PR, check CI, merge | | Coordinate | Squads | Set an objective and a team, then accept or discard | | Repeat | Automations and squad schedules | Run a prompt or a saved squad on a schedule | _The multi-agent coding workflow, mapped to the app_ ## How do squads orchestrate a team of agents? A **squad** is a leader agent plus member agents that Tallos runs as one job. You write the objective; the leader splits it, starts each member in its own worktree, checks what comes back and repeats. One round is plan, dispatch, review. - **Templates:** Build + review; 3 attempts, pick the best; Leader divides the work; or Custom. - **Team:** every member has its own agent, model, effort level and job; **How many** runs several copies of one member. - **Limits:** 5 rounds and 4 members at once by default, up to 20 rounds and 12 members. Tallos refuses members over the limit; the leader stops and reports at the round limit. - **Questions:** when the leader is truly blocked, it asks with `tallos squad ask` and waits. You pick an option or write your own answer. - **Final review:** the leader reports with `tallos squad report`; you see results by worktree, one marked **Recommended**. The leader is instructed never to merge or push. You **Accept** or **Discard** the run, and accepting does not merge either: you open the recommended worktree to create a PR or merge. Squads can be saved, reused and scheduled. See [squads](https://runtallos.com/features/squads). ## How do review, accept and merge work? Review is the gate between an agent's work and your codebase. Every workspace has a diff against the branch it started from, inline or side by side. Leave notes on any line, then use **Send notes to** to hand them to the active agent session or a new agent: one prompt, every note anchored to its line. When the diff is clean, Source Control walks you through **Commit**, **Push** and **Create PR** on GitHub, GitLab or Bitbucket Cloud, and linked GitHub pull requests show checks and reviews in the app. More in [diff review](https://runtallos.com/features/diff-review). > **Nothing merges on its own** — Tallos never merges an agent's work by itself. It stays on the workspace branch until you merge it or delete the workspace. ## Can agents run on a schedule? Yes. An **automation** is a prompt an agent runs on a schedule: hourly, daily, on weekdays, weekly or on a custom cron expression. Each run can start in a fresh workspace or continue in an existing one. Saved squads can run every hour, every day, on weekdays or every week. Results wait for your review. See [automations](https://runtallos.com/features/automations). ## How is this different from running agents in terminals by hand? You can build the same workflow by hand with terminals and `git worktree`. It works for two agents; the overhead grows with each one you add. **By hand: terminals + worktrees** - `git worktree add` and a branch name for every task - One terminal window or tmux pane per agent - Check every pane to see who is waiting on you - Read `git diff`, then retype your feedback into the prompt - Split and hand out the work yourself - Push, open the PR in the browser, remove worktrees by hand **In Tallos** - **New workspace** creates the worktree and branch in the background - Every agent in its own pane, as chat or terminal - Status icons and notifications show who needs you - Line notes on the diff go back to the agent in one batch - A squad leader splits the work and checks the results - Commit, push, PR and cleanup from the same window > **Still plain git** — Every Tallos workspace is a real git worktree. Open a terminal in it and use git as usual; Tallos picks up the changes. > Stop wiring terminals together. Run the whole loop in Tallos. → https://runtallos.com/signup ## Frequently asked questions ### How does an ADE work? An ADE gives each task an isolated workspace, runs an AI coding agent in it and puts the resulting diff through your review before it merges. Tallos does this with one git worktree per task and the agent CLIs you already use. ### What is a multi-agent coding workflow? It is a workflow where several AI agents work on separate tasks, or the same task, at once, each in its own workspace, while you plan and review. In Tallos you run it with parallel workspaces or a squad. ### Does Tallos merge code automatically? No. Work stays on the workspace branch until you commit, push and merge. Accepting a squad run does not merge either; you open the recommended worktree to create a PR or merge it. ### Do agents in Tallos run commands without asking? Not by default. Tallos starts agents in manual mode, so they ask before running commands. You can opt into autonomous mode under Advanced on the welcome screen or in Settings → Agents → Agent Permissions. ### What is AI agent orchestration? AI agent orchestration is when one agent plans the work and coordinates others. In Tallos, a squad leader splits the objective, starts members in their own worktrees, checks results and asks you only when blocked. ### What happens to the workspaces I don't keep? Delete them. Tallos asks for confirmation, then removes the worktree folder and its branch. When you discard a squad run, you can also delete the members' worktrees. ## Sources 1. [git-worktree documentation](https://git-scm.com/docs/git-worktree) — Git 2. [Run parallel sessions with worktrees](https://code.claude.com/docs/en/worktrees) — Anthropic 3. [What is the Model Context Protocol (MCP)?](https://modelcontextprotocol.io/docs/getting-started/intro) — Model Context Protocol --- **Run your agents in parallel with Tallos** — Claude Code, Codex, Gemini and 30+ agents — each in its own workspace, with squads, chat, terminals and review in one app. Uses the subscriptions you already have. https://runtallos.com/signup (macOS 13+ · Windows 10+)