My Workplace: a team chat of AI teammates (starter kit) - #36
Merged
Merged
Conversation
…ct messages, groups with turn-taking and @mentions, artifacts) A team chat where every teammate is an AI agent: a recruiter Harness interviews the person (JSON questionnaire rendered as multiple choice) and the app installs the teammate it designs as its own Harness with the teammate package. Direct messages are one session per person per teammate; a group is one session on the recruiter whose group.json holds the roster, the transcript and each teammate's cursor, with the posting tab running the replies: mentioned teammates first, everyone else judges relevance with [skip], replies that @mention pull others in up to three hops. The browser a teammate opens shows live in the rail; files are artifacts per conversation and in one library. Twelve drawn faces ship with the kit. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…om, a direct message keeps its name after the first turn, artifacts settle and refresh, the rail closes on a narrow window, the library stops reloading on every poll Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…s finish renames it after the message it answered Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…he moment the page loaded Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…nd the example pills start the interview; a slow recruiter says so Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…, a soft grey room list with previews, black and grey bubbles, a pill composer, white cards in the panel, a face-first hiring screen); the app scopes its calls to the console's active workspace and picks that workspace's recruiter; one answer is enough to continue the interview Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…rary is a grid of previews with makers and opens the shared viewer in a sheet; the questionnaire keeps an even rhythm; an open conversation reads the server every three seconds so a message from another window shows at once; reifyui 0.13.5 gives an HTML artifact a Preview | Source switch Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…er's roles and files, and a fresh workspace gets its browser The group's turn-taking is derived from the document (who has read what, who is typing, who holds the round) instead of the posting tab's memory: a tab that reloads or closes mid-turn hands the room to the next one after a minute, which reads back the reply the vanished tab never wrote (the turn ran on in the teammate's session) and continues with whoever still owes an answer. A refresh had lost a finished launch plan and nobody followed it up. Watching windows follow a running turn on the Harness event stream, keep the bubble until the reply is in the room, and the driver keeps the typing mark fresh for long turns. Every turn carries the roster with roles and expertise and the rule to hand work to the role it fits; files other teammates shared arrive under shared/ and the library lists them once. Hiring connects the browser plug for a workspace that lacks it. Spreadsheets preview in the viewer. Artifacts page has a search box; rails show the same cards; the browser card keeps a minimum height. DeepSeek harness + Claude Sonnet 5.5 is the first recommended pairing. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ed for again; README describes the resumable room Sonnet 5.5 on the DeepSeek harness once closed the profile object with a dangling comma and quote; the page showed 'cannot draw' for a valid profile. The parser repairs that tail, and a reply that still does not parse is asked for once more before the person sees anything. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…te is refused at the reveal A fourth hire came back as a second Marlowe: two teammates with one name share one @handle. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…and the library Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
A team chat where every teammate is an AI agent with its own face and its own job. A recruiter Harness interviews you (JSON questionnaires rendered as multiple choice) and the app installs the teammate it designs as a Harness of its own. Message a teammate one to one in a conversation it remembers, or put several in a group where they answer when it is their turn, hand work to each other with @mentions, browse the web in front of you, and leave what they make in a shared library.
How a group answers. The group document (
group.jsonin a recruiter session) says who still owes the room a turn: the mentioned first, then everyone else judges with[skip], a reply's mentions pull others in up to three hops. Whichever window has the group open runs those turns; a window that closes or reloads mid-turn hands the room over after a minute, and the next one reads the finished reply back from the teammate's session and continues. Every turn carries the roster with each teammate's role and expertise, so work goes to the role it fits; files colleagues shared arrive undershared/in the teammate's working directory. Other windows follow a running turn on the Harness event stream, so the teammate's tools and text show everywhere as they happen.Recommended pairing: DeepSeek Harness with Claude Sonnet 5.5 (then Claude Code with Sonnet 5.5, Codex with gpt-5.4, Hermes with deepseek-v4-pro). A fresh workspace gets its browser plug connected at the first hire.
Verified end to end on hr-test with headless Playwright: a new workspace, launch from the Kits page, four teammates hired through the questionnaire, one group, a launch campaign for a client run in it (brief, LinkedIn post and email, two PNG images, X thread and posting schedule, a review pack) in 8 minutes on DeepSeek Harness + Sonnet 5.5; a window reloaded mid-turn recovered the reply and ran the follow-ups; a second window showed the turns live; spreadsheets, Markdown, code and HTML preview in the pane; four widths. The gateway fix the first round found landed separately (HarnessRouter/harnessrouter#347); the card's cover is HarnessRouter/harnessrouter#348.
🤖 Generated with Claude Code