Repository navigation
Releases
Agent Workbench v1.0.0-beta.3
The third pre-release of the next major. It closes what beta.2 listed as known and what turned up since.
An integrated session starts in the mode you set. A conversation started in the Default permission mode whatever the agent's own settings said, so it asked where a terminal session would not. It takes the agent's settings now, and a session another session starts does too. A row can be reopened the other way, as a conversation or in a terminal, and keeps the mode its chip last showed.
Send now, and take back. A message written while the agent works waits under the thread. The line under the last waiting message sends them now, and each waiting message of your own can be taken back into the composer. Claude Code takes the message into the running turn where it can and stops the turn for it otherwise; Codex steers it into the running turn. The waiting line stays above the composer.
A long conversation reaches its foot. In a conversation of several hundred turns the view could stop short of the last line, and switching between long conversations stalled the window. The foot is reachable by wheel, keys and the scroll bar, and a conversation that is not on screen costs no layout.
Settings and sliders answer at once. Dragging the Cards control or a colour repainted every terminal on each step. Terminals are recoloured only when their colours change, and opening the settings no longer stalls.
Usage figures stay fresh. The status bar's usage is fetched for each account in use about every five minutes and when a session of an account without figures comes to the front, and the last figures are kept across a restart with their age.
A session that starts others hears the truth. A started session was reported as stopped during any quiet spell of a long turn. A turn counts as ended only when the agent says so or you stop it. A turn an error ended is said for what it is, a usage limit with the time it lifts; a stop is told once; the list says which sessions are closed and which can be resumed; and a name given at start stays through a close and a resume.
Also. In a Codex session, in a terminal or integrated, ten of the workbench's own tools run without asking: the ones that only read, and the ones whose changes the workbench asks you about itself. Every other tool, the browser and the ones that start, stop or write to sessions among them, follows the session's approval policy. The background's gradient is drawn finely once it rests, without stair-steps. The sessions list's scroll bar lies over the pinned project headers.
Installing. brew upgrade on the beta channel, or brew install --cask agent-workbench@beta from the elva-labs/elva tap after brew uninstall --cask agent-workbench. A paired machine needs the daemon at the same version: brew upgrade agent-workbench-remote-beta there, or the install script with AGENT_WORKBENCH_VERSION=v1.0.0-beta.3.
Agent Workbench v1.0.0-beta.2
The second pre-release of the next major. It fixes the freezes of the first.
The window froze with a large changes tree. Each row of the file tree held an ordering number of its own, and the macOS webview walks the list once for every distinct number on each layout and paint, so a tree of a few thousand rows made any layout take seconds: typing in a terminal, opening the tree, opening settings. The rows hold no number now, and the same layout takes a few milliseconds.
See-through cards show. In dark the tinted ground was painted lighter than the cards, so the two ran together, and a small share moved a card by a level or two. The ground is darker with more colour, a card takes on the background's colour and keeps its own lightness, and the Cards control runs from 0 to 100%. A share saved in beta.1 reads four times higher on the new scale and looks stronger.
The conversation spans its pane. An integrated session's thread, composer and dock use the pane's width; code, diffs, output and tables take all of it, and prose wraps at a comfortable measure.
Colours that belong together. Adding a third colour to a background places the others close to the first, about 50 degrees either side, and the first, drawn larger, carries them as it is dragged. The ready-made backgrounds are retuned. A background saved before is drawn as it was.
A read session stays read. A session's row went unread again a minute after it stopped, and again after each recap or compaction Claude Code did while idle, since its idle notice was taken for news. It is marked now only for something new.
Also. The focused sessions pane carries its mark under its header, as the other panes carry theirs along their cards. The orchestrator project is retired: to direct work, pick a workflow, among them a built-in Direct, or ask any session to start others; a folder of old orchestrator sessions is listed as an ordinary project while it has history. Two faults found by running Codex against a real model are fixed: a failed turn was told twice, and a session named no effort when Codex set none.
Known in this pre-release. An integrated session starts in the Default permission mode whatever the agent's own settings say, so it asks where a terminal session would not; a long conversation can stop short of its foot when scrolled; a message written while the agent works cannot be sent at once. Fixes for all three are in review.
Installing. brew upgrade on the beta channel, or brew install --cask agent-workbench@beta from the elva-labs/elva tap after brew uninstall --cask agent-workbench. A paired machine needs the daemon at the same version: brew upgrade agent-workbench-remote-beta there, or the install script with AGENT_WORKBENCH_VERSION=v1.0.0-beta.2.
Agent Workbench v1.0.0-beta.1
The first pre-release of the next major. It runs on the settings and state of 0.8.0 and hands them back to it, so moving between the two loses nothing.
Sessions as conversations. An agent can run integrated: Claude Code and Codex are driven through their own structured protocols and the conversation is drawn natively, with your words in a bubble, the agent's steps folded into one line, permission requests docked over the composer, and your own login, skills, hooks and MCP servers as before. A session in a terminal stays a choice in Settings. A resumed conversation shows its earlier messages. The composer completes slash commands and skills with /, mentions a file with @, and attaches any file. Cmd F finds in the conversation, each part can be copied, and a long conversation stays fast.
Workflows. A workflow gives sessions fixed roles, an orchestrator with work and review steps, drawn and edited as a diagram. A new session picks one from a chip in the composer. The header shows where a run stands, started sessions hang under their starter with their subagents, a starter answers its sessions' requests from its own dock, any session that started others gets a board at the foot of the changes pane, and an idle starter is woken when a session it started stops or asks.
A new frame. Panes stand as cards on the window's ground with their headers in one band, and a divider is drawn only while it is dragged. The sessions pane splits active from inactive by itself and sorts by last used or name. Cmd P goes to a session or project and Shift Cmd P runs a command.
A background of your own. Appearance takes a colour or a gradient behind the chrome, with strength and grain, and the cards can let a share of it through. Text keeps its contrast whatever is chosen.
Around the edges. Shells and the processes list belong to the session they were opened beside. Cmd T opens a shell, Cmd W closes what is in front, and closing the window or quitting asks while sessions or a busy shell are running. Editor mode takes untitled files, named when first saved. Browser tabs belong to the session that opened them, and an agent can take a screenshot of its own tab while you are elsewhere.
Installing. brew install --cask agent-workbench@beta from the elva-labs/elva tap, after brew uninstall --cask agent-workbench if the stable app is installed; the two are the same app and cannot sit side by side. brew upgrade then follows the beta. A paired machine needs the daemon at the same version: brew install elva-labs/elva/agent-workbench-remote-beta there, or the install script with AGENT_WORKBENCH_VERSION=v1.0.0-beta.1. Windows gets the NSIS installer and Linux the .AppImage and the .deb; a pre-release has no .msi and no .rpm.
Known. A custom stylesheet written for 0.8 may need changes for the new frame. Going back to 0.8.0 with the .deb is a downgrade to dpkg and has to be asked for as one.
Agent Workbench v0.8.0
A command palette. Shift Cmd P (Cmd ; in the vim preset) puts a field over the workbench listing every session and every action the keys can do. A session is listed in the order the list by status has it, with its project, host, account, the orchestrator that started it and the line it left; an action is listed with its chord. What is typed narrows the list by every word, in any order. Up and Down move, Enter picks, Escape or the chord again puts it away and the keyboard goes back where it was. While the palette is up no other chord gets past it.
The next session that needs you. Cmd . goes to the next session waiting on you and puts the keyboard in it: those asking for a permission first, then the one that has waited longest, stepping round from the session on screen so pressing it again visits every one. With nothing else waiting the status bar says so.
Zoom per terminal. Cmd =, Cmd - and Cmd 0 make the terminal with the keyboard a pixel larger or smaller, 8 to 24, or 12 again: the panel's shell when the panel has the keyboard, the agent otherwise. Each terminal has its own size, and an agent session's comes back after a resume or a restart. The grid is measured again and the process told, as for a font change.
All five are ordinary keymap entries, rebindable in Settings → Keys. A keymap saved before them takes the preset's chords for them, never one the keymap already gives to something else. Switching sessions with the keyboard in a terminal no longer leaves the keyboard on the body.
Upgrading. The desktop only pairs with a daemon at its own version, so a paired machine needs 0.8.0 as well: brew upgrade agent-workbench-remote there, or the install script again. On macOS, brew upgrade --cask agent-workbench from the elva-labs/elva tap.
Agent Workbench v0.7.9
style: format the delegated stop test
Agent Workbench v0.7.8
chore: bump to 0.7.8
Agent Workbench v0.7.7
chore: bump to 0.7.7
Agent Workbench v0.7.6
chore: bump to 0.7.6
Agent Workbench v0.7.3
test(driver): key the account override by the path the window holds The account test wrote its override under the temporary repository's path as Node spelled it and expected the project's row to show the account. The window holds a project under its resolved path, so on Windows the runner's short temporary directory never matched, and on Linux the remote test before it had left only the ssh project open. The test now opens the local repository itself and keys the override by its resolved path.
Agent Workbench v0.7.1
The whole answer on screen. The agent was told to call notify as the last step of every turn. Claude Code folds what an agent writes before a tool call into one line, so following that instruction folded the answer away and left only the sign-off the agent added after the call. The agent is now told to call notify just before its final message and to write nothing after it, so the answer is the last thing in the turn and shows in full. The same goes for the line a session started by an orchestrator is asked to leave.
One build per update. Updating a plugin from its repository stops its process and starts the next one, but the old process's reader scheduled a restart of its own as well, and the two starts could both run the plugin's build. A process the workbench stopped is no longer restarted by its reader, and starts of one plugin happen one at a time.
Upgrading. The desktop only pairs with a daemon at its own version, so a paired machine needs 0.7.1 as well: brew upgrade agent-workbench-remote there, or the install script again. On macOS, brew upgrade --cask agent-workbench from the elva-labs/elva tap.