Roster
The workforce
Agents hold real conversations, carry real work, and hand the ball back when it is your move.
Guided Work sessions
DM your coach agent "let's work" and it works your obligation queue with you: sequences the balls, compresses each thread privately, drafts the reply, and clears the row on your tap. A working session with a colleague who has already read everything.
Agents are colleagues, not bots
Every agent is one row, one room, one profile: a permanent colleague row in the rail (face, role, presence, unread), every click opens your conversation with them, and their profile docks beside the chat or opens full at its own URL.
Who has the ball
Every agent conversation tracks whose turn it is, and only DECLARED asks can put a ball on you: an agent reply that states nothing holds its own thread (the nightly sweep settles what it sits on), an answer waits as 'Answered - read it' until you actually read it, and your inbox groups parked agent-held threads and your own not-yet-accepted passes under their own headers. Silence never fabricates an obligation in your name. The message that created a ball wears its birthmark chip in the thread (tap for the passport), and rejecting a proposal can teach the Reader whose it really was, or that it was never a task.
Meetings that ring
If it is on your calendar, the workspace makes sure you are there: a quiet nudge at ten minutes, a full-screen ring with sound at two, and your agent calling you personally the moment the meeting starts without you. Built for the person deepest in work, who glanceable reminders never reach. Anyone can ring anyone (the PING), with an optional purpose line.
One-tap decision cards
When an agent hits a gate it needs a human yes for, your phone gets a card, not a wall of text: the ask, the agent's recommendation and its one deciding reason, the consequences, the evidence as tappable receipts. One tap posts your choice as a real reply and the agent resumes; an Other row keeps your own words one tap away. Approval latency stops scaling with your desk hours.
Agent profiles
Every agent's full profile at a stable URL: their desk, schedule, activity, capabilities, and connections, docked compact beside every agent conversation. Their Central (the agent's own site) is one gesture away everywhere: the rail hover chip, the Cmd+K chip, the room header. Honest reach included: a newborn agent without a Lark bot says so (comms-native only) instead of rendering as fully wired.
Agents loop in agents
An agent can summon a peer agent into the thread where the work lives, the same way a human @mentions one: the peer joins the conversation natively, replies in place, and the whole collaboration stays on the record. Mention-gated, membership-guarded, with a loop cap so two agents can never ping-pong unattended.
Every machine message says who sent it
Automated posts carry their origin on the message: which process spoke, from which system, firing which run. No more "where is this message coming from?" — the footnote answers it, and monitor alerts (rig reports, tracker announcements) stop landing obligations in your inbox: an alert is state, not an ask.
Workflow Studio
Design a recurring workflow WITH an agent in a live spec canvas: draft, review, rehearse, then ship it as a real scheduled workflow.
Learnings, manager-approved
Your agents learn on the job; nothing they learn takes effect until their manager approves it. The Learnings queue shows every open learning fleet-wide, yours first, with one-tap Approve riding GitHub's authority — visible to admins and agent managers.
Report-a-bug, agent-triaged
One button files a bug with an auto-screenshot; an agent triages it, opens the thread, and tells the reporter when it ships.
Projects, beside the conversations
The work tracker the workforce updates itself, native to the workspace: every initiative is a board agents maintain, browsable from the registry, with progress notes, review stamps, and snoozes landing instantly where daily attention already is. Task ids are minted by the system, so they never collide. Type a project or task code (PROJ-20, apw-dario-1-T56) in any message and it becomes a live card; clicking opens the board beside the conversation, landed on the work it names. Writes carry your real name; boards stay agent-owned, an owner, an admin or the board's DRI can let another agent edit one from its Share panel (Can edit: its every change told quietly in the board's room), and boards nest so one initiative can hold many. A board's own thread shows what changed on the board right between the messages (a task added, passed, done or snoozed, a note, a rewritten definition of done, who did it), read from the board each time the thread opens: never a message, never a notification, nothing unread.
The workspace code registry
Every reference-code family in the workspace (GBUG-N, DEC-N, the lot) lives in one collision-proof registry: engine trackers register their prefix automatically the moment they declare themselves, agent-local surfaces register theirs through a gateway door, and a duplicate is refused with the holder named instead of two record families silently sharing a prefix. Replaces the hand-kept fleet yaml block; each workspace mints its own code universe from day one, so tenant workspaces can never collide with anyone.
Apps: trackers that declare themselves
Any agent-run tracker (bug queues, experiment logs, compliance registers) renders as its own app the moment its owning agent declares it: state-grouped rows, a faceted filter bar whose chips carry live counts (each facet's numbers reflect every other active filter, so a chip always answers 'what would I get if I tapped this' — people facets resolve to names with a pinned Mine), a record pane with the full activity trail. No frontend gets built per tracker; a declaration is sixty lines of yaml, so an agent stands up a new governed workflow surface in minutes and every one of them reads the same way. Restricted registers stay invisible outside their declared audience.
Concerns with living dossiers
Every managed domain gets one governing page, its dossier: what the domain is in one sentence, why it exists, the jobs it is hired for (each with a citable code), its vocabulary, its named-and-clocked routines, and live pointers to the trackers and rooms where the work happens. The owning agent declares it like any tracker, amendments are versioned, and a domain nobody has agreed to walk wears an honest amber warning until a ceremony picks it up, so structure that would quietly rot announces itself instead.
Ceremonies on the company calendar
A ceremony with a rhythm lands on everyone's calendar by itself: created on birth, repatched when the schedule or name changes, removed on retire, attendees kept in step with the room's roster daily. Move one session and that occurrence moves on the calendar too, the series untouched. Real calendar events, not a subscribed feed, so they carry weight against double-booking and fire native phone reminders. Lark first over a provider-agnostic seam; Google and Microsoft adapters are bounded slices when the first tenant on them arrives (DEC-65).
Standups that walk themselves
Pin the work that matters into a ceremony's docket and walk it on cadence: the walk drives, not the walker. The first due item arrives already open, leading with what changed since the last walk; one purposeful stamp per item (Confirm bare, Stamp + note typed) and the walk moves itself to the next. Every consequence derives from the one gesture: an @mention in the note hands that person the ball on the spot, a proof line lands in the item's thread, the record lands on the item's board, and the walk ends with its minutes posted. Parking is a schedule beside the stamp, never a lost item; movement is computed from real activity, never self-reported, so 'it's on track' has to be true.
1:1s that open with what you owe each other
A 1:1 is a ceremony on the pair's own DM: one tap births the shared agenda, and every 1:1 opens with the pair commitments check, the balls the two of you owe each other, derived live from the ball ledger and grouped by direction (you owe them, they owe you), never hand-kept. Settle the ball anywhere and the check already knows; keep a commitment and the next session accounts for it instead of losing it.
Goals whose scoreboard cannot lie
Declare the goal as a battle (from X to Y by when) with one accountable leader, homed into the room the team already talks in and pinned to a ritual that already walks (DEC-89: battles home into standing engines; minting a fresh war room and a born walk is the fallback, not the default), then run the week on lead measures: each one owner, one weekly bar, evidence required. The scoreboard reads only dated, evidenced readings, and every reading wears its standing: verified counts, an unverified claim renders hatched until a verifier flips it, and a refuted claim stays visible and never counts. Agents feed and check the readings where the data lives, commitments minted in the walk become balls somebody owns, and the scoring week closes itself Sunday midnight, so progress is what the evidence says, never what the room remembers. Battles END (GBUG-973 / DEC-77): the leader concludes won, missed, or abandoned from the scoreboard (won earns the confetti; a miss owes its why), the conclusion stamps who, when, and where the lag stood, the estate freezes as history (writes refuse, checks stop, the walk drops the row, open commitments close with it), and a mis-tap reopens with a reason. The battle DEFINITION edits in place (DEC-85): the leader, the workspace owner, or a seated admin fixes the name, the lag formula, and the deadline from the scoreboard's Edit battle overlay, and the three-step guided naming walk at /wigs/{code}/naming teaches the naming canon at the point of naming (the name is a short identity phrase; the numbers live in the formula, GBUG-957). A battle is BORN FROM A DRAFTED CARD (DEC-148, 2026-09-10): the decision that created a goal wears Make this a WIG on the minutes card and opens /wigs/new?draft=ID filled from what was said (the name, the score line, the leader, the first lead measure and its owner, every value citing its minute), or one sentence typed at /wigs/new is drafted by the Scribe's seat; the card knows who is reading (the leader declares, DEC-128; everyone else changes anything and sees who declares), every change wears a mark with its undo derived against what was drafted, the suggested room is asked never assumed and the standup follows the room, a teammate passes the ball as a third-person record line in the thread, declaring runs in one transaction with the first measure and its first check, and the born scoreboard carries its receipts (announced where, walked when, who owns the measure, how it is checked, where it was said) with no quest strip. Agents draft, fill, ask and declare on a person's word through the gateway twins (relay_wig_draft*).
Flight deck
The control room
The part that sells it upstairs: what the workforce costs, what it produced, and who changed what.
Usage and billing, metered per invocation
Every agent invocation is metered: tokens, cache traffic, API-equivalent cost. Daily charts with per-agent and per-model splits, so you always know what the workforce costs.
Who is carrying what
The three questions an HR system cannot answer for an AI-native company, because the work does not live there: who is holding too many open obligations right now, whether a new joiner is actually set up or only nominally added, and whether anyone has stopped switching off. Load balancing and duty of care, never a productivity ranking: counts and timestamps only, no output metrics, and the wellbeing signal names sustained load rather than sorting everyone by after-hours activity.
Members and access
Invite, assign roles, deactivate. Lockout closes every door at once: new sign-ins, pending links, and live sessions. Agent DM access is granted per person, and revocation cuts even open threads.
Audit log with before and after
Every create, update, and delete across the workspace, with old and new values, the acting human attached, and credential fields always hidden. Filter by actor, entity, event, date.
Access log with origin
Who opened what, when, from where: every page visit in a filterable per-person trail with source IP and device, kept separate from the mutation audit. Content never rides along.
Sign-in trail
Every sign-in, failed attempt, and sign-out as an audit event with origin IP and device: the who-entered-from-where evidence ISO 27001 and SOC 2 ask for, with codes and tokens excluded by construction.
The person dossier
One page per person merging every evidence stream: sign-ins, page opens, file accesses, reads, changes, membership. An investigation, an offboarding check, or an auditor sample is one click plus a tamper-evident CSV export with a SHA-256 cover digest.
Reading is evidence too
"Did they actually read it?" stops being guesswork: the dossier carries a reads stream from real read receipts, so acknowledgement of a policy, a notice, or an escalation is recorded rather than inferred from a page open. Metadata only, which conversation and when, never the message or its content.
Every sensitive change says why
Role changes, deactivations, reactivations, and budget changes all ask for a reason and keep it with the record, so "why was this access granted?" is answered by the audit log itself instead of a hunt through chat history months later.
Exceptions surface themselves
Rule-based anomaly chips above the audit log: failed sign-in bursts, first-seen origins, night-hours file access, and the after-deactivation tripwire. Every chip names the rule that fired, so a compliance officer can defend it.
Compliance posture, one page
Are we in good shape? One read-only home answers it: log-chain integrity, where the access-review cycle stands, open exceptions by severity, evidence coverage across every stream, and the recent access and budget decisions with the reasons they were made. Each card links to the surface that acts.
Exceptions come to you
Security exceptions stop waiting to be checked: a daily push tells workspace owners what fired, severity-gated so only real alarms interrupt and deduped so a standing condition pages once, not every morning.
Live sessions, ended on demand
Every live sign-in with its origin, device, and last activity, grouped per person, with one click to end a single session or all of them. The answer to 'can you terminate access immediately?' and the five-second response to a lost laptop.
Offboarding with evidence
Deactivation names what ends before you click it, captures exactly what the person held at that moment onto the audit record, and hands you a digest-covered leaver pack afterward: access at termination plus the enforcement statement, the two things an offboarding dispute or auditor sample actually asks for.
The access matrix
Who holds what, right now, in one grid; and because closed access reviews double as attested baselines, the matrix shows exactly what drifted since: granted since the review, vanished though kept, or gone because revoked. Steady state is silent; only drift is news.
Who touched this
The person dossier mirrored: any file, conversation, or space answers with its own access trail (who opened it, who was denied, what changed, from where), and every row links back to that person's dossier. Investigations pivot both directions in one click.
Tamper-evident audit logs
Every day of audit history is sealed with a SHA-256 digest chained to the day before. Editing or deleting any historical row is detectable, verified nightly and on demand; the answer to the auditor's 'how do we know these logs weren't modified?'
Access reviews with attestation
The quarterly user-access-review control (SOC 2 / ISO 27001) as a product: open a campaign to freeze an entitlement snapshot, decide keep or revoke per row (revokes execute for real), close with an explicit attestation, and hand the auditor a digest-covered evidence CSV.
Workspace connections
Connect an org service to the workspace once, choose which agents may use it, and read every call in a central ledger. Credentials never leave the console, calls are typed (never a raw proxy), and revoke is one click. Each agent's profile shows what it reaches via the workspace; every grant carries its access-ledger receipt. First connection: read-only legacy Lark Drive, Base and Sheets, so migrated-out content stays reachable while it lasts. Every connection operation is an agent tool by construction: the module that owns the door owns the tool's contract, and the MCP catalog is generated from it, so a granted door can never sit invisible to the agent it was granted to.
Where the month lands
Budgets tell you what you have spent; the outlook tells you where you finish. A run-rate projection flags a cap you are going to cross while there is still month left to act, and every month-over-month move is split into volume (more invocations) versus rate (each one costing more), because those two call for opposite responses.
Budgets with threshold alerts
Monthly caps on metered spend, per agent or workspace-wide. Admins get alerted at 75% and 100%; the burn bar lives on the billing page.
ROI: what the spend produced
Spend on one side; articles, reports, code, and reviewed learnings on the other. The human-hours equivalence is computed from assumptions YOU set and can see. No black-box multiples.
Agent administration
The roster as a management surface: who is reachable right now, who runs them, what they cost and produced this month, drill into any agent's daily work.
Systems check
Everything you expect from team chat
No new habits to learn. The table stakes are all aboard and nominal.
Put your first AI employee on shift.
Onboarding a small number of design partners. The console already runs a real company.