The Orchestrator (Orbi)

The Orbi orchestrator is the special agent created automatically in every workspace. It does not do the work itself, it hands the work out. The platform is called Orbi too, so we always name the agent in full. "Orbi" on its own means the product, while "the Orbi orchestrator" means the agent that hands out the cards.

When you assign a task to the Orbi orchestrator, four things happen.

1. It analyses the task and splits it into steps.
2. It hands each step to a suitable agent.
3. It judges the result of each step and revises the plan if needed.
4. When the work is finished, it moves the task to Review.

#The plan-and-execute pattern

The orchestrator draws up a rough plan at the start, but it does not lock itself to that plan. After every step it asks "does this result change the plan?" and adds, removes or reorders steps when the answer is yes. A single task can run through several cycles.

flowchart TB
  Q["Task queued"] --> P["Cycle 1: draw up a plan"]
  P --> CL["Checklist steps"]
  CL --> E["Run the step<br/>a suitable agent"]
  E --> V{"Is the plan<br/>still valid?"}
  V -->|"yes"| N{"Any steps left?"}
  V -->|"no"| RP["New cycle:<br/>revise the plan"]
  RP --> CL
  N -->|"yes"| E
  N -->|"no"| R["Review"]

This matters because real work cannot be planned in full up front. On an "analyse the competition" task, an unexpected segment may turn up while the third competitor is being researched, and a fixed plan would have ignored it.

#Checklist state machine

Sub-tasks are not separate cards, they sit inside the main card as checklist steps. Every step has a state, and that state is pending, in progress, done or error.

Three guarantees hold.

• Order. A step does not start before the step in front of it has finished.
• Resume. After an error or an interruption, "where did I leave off?" is read from the checklist, and a step that is already done is not repeated.
• Steps are never deleted. A failed step is marked as an error rather than removed. The reason is straightforward, because a deleted step also deletes the answer to "why did it fail".

Every step carries the cycle it was completed in, so you can read from the card's history how the work developed.

#The agent works until the work is done

When a cycle ends with checklist steps still open, the card does not drop to review; the agent starts the next cycle right away in the same run and the card stays "running". The chain stops in three places: when the checklist is closed, when the cycle budget is reached, or when a cycle closed no step and added no output. In that last case the card goes to a person and the summary says why, because repeating a cycle that made no progress only spends credit.

Before a cycle closes, the agent reads its own work against the request once: is every item delivered, in the number and shape asked for, was anything written twice, is a failed tool call still unfinished. If something is missing it completes it; otherwise it closes. This turn does not run on a plain question-and-answer turn without tools.

Between cycles the agent leaves itself a working note: sources read, decisions taken, the next step. The note is shown on the card's Summary tab, only the agent writes it, and it is not a deliverable; deliverables live in Outputs and communication in comments.

Every cycle sees the list of the card's outputs, not their text. If the agent writes its report again in the same run, the new text does not become a second report on the card: it becomes a new version of the first one and stays one document in the Outputs tab. When the card asks for more than one deliverable (a blog post and a LinkedIn post, say), the agent adds the second as a separate deliverable.

The Runs list on the card's Trail tab shows, row by row, who worked on this card, where and on which branch: Orbi's own run, a run taken by a device and a delegate working inside a run are separate rows. A card can carry several runs; each row shows its state (running, finished, released, error) and its duration.

#Ephemeral agents

If none of your existing agents fits a step, the orchestrator can create a temporary agent for that step and give it its own system prompt. This agent is not added to your workspace's agent list, and it lives only for that step.

#How review feedback works

The feedback you write in Review does not go to an agent directly, it goes to the orchestrator first. It reads the feedback and decides which agent or agents run again. A comment like "the third section is weak" therefore repeats the step it concerns instead of redoing the whole job.

#Budget ceilings and loop protection

Where an agent that replans itself will stop is not knowable up front. Orbi therefore puts two ceilings on every task.

• The maximum number of cycles (orchestrator revisions), 10 by default
• The maximum number of total agent runs, 30 by default

The value applied is the lower of the system-wide fixed ceiling and your plan's limit.

If a ceiling is passed, the task drops to Error on its own and an explanatory comment is left. Nothing is lost at that point. The checklist stays, completed steps are kept, and requeueing carries the task on from where it stopped.

An agent cannot see its own limit. This is deliberate, because a model that knows the limit tends to leave the work half done and lower the quality when it thinks it has to hurry. The limit stays entirely at the system level.

#Proactive mode: let Orbi act like a teammate

By default Orbi only works when you ask it to. With proactive mode on it becomes a teammate who speaks up on their own.

• Morning briefing. It leaves a short message in the channel on the days and at the time you pick, covering what finished yesterday, who is waiting on whom, what stopped and what is on today. When there is nothing to say, it stays quiet.
• Reminders. It leaves a single comment to the owner of a card that has sat in review too long, gone silent or has nobody on it, and it does not write to the same card twice in one day. The level (Gentle, Normal or Insistent) sets how long it waits and in what tone it speaks.
• Weekly plan. It brings a summary of the week and the cards to pull forward for the next one. This is a proposal, not an order. Orbi never changes a card's state, assignee or priority on its own.

While doing this, Orbi does not merely count the board, it reads it. Every day it goes through the last seven days of cards, comments and channel talk and works out which subjects the team is actually working on. For each subject it finds similar work done in the past and related cards forgotten in the backlog. Instead of listing events, the morning briefing talks about these subjects, so it says what a subject is, where it is heading, what was done before and what Orbi thinks (its own view is marked as such). If the subject is an outside technology or market development, Orbi looks it up once on the first day and does not open a card for it. Reminders follow the subject too. A card that belongs to what the team is discussing today but sits forgotten in the backlog gets a note to its owner, while unrelated old cards are left alone. You can see how Orbi reads the team in the Agenda panel on the same page. If the reading looks wrong, make the cards and the channel clearer and the next reading corrects itself ("Read now" refreshes it without waiting).

#Evidence-backed contributions

Orbi joins the conversation not only by summarising it but by speaking up when it has evidence in hand. This takes four forms: it points out two statements on the board that contradict each other, quoting both; when the team discusses something that needs outside research, it looks it up itself and brings the answer with its source; it proposes a concrete way Orbi could do the thing being discussed; and, rarely, it offers a genuinely new idea. A contradiction claim without its quotes or research without its source never gets written — that requirement is not a setting but the system itself. The nudge level also decides which contributions arrive: Gentle brings only contradictions and research, Normal adds feasibility notes, Insistent opens up ideas too, and their number stays under the daily cap. Every contribution shows up in the Recent notes list on the same page. A kind of contribution that keeps getting no response is scaled back automatically, and saying Don't suggest again writes the lesson to Orbi's memory.

#Work Orbi hands over and the outside-world boundary

Orbi can also hand the work of a live subject to a suitable team agent by itself: it opens a card, writes the context and the reason into it, starts the agent right away and tells you about it. This work has one boundary: nothing Orbi started on its own can send anything to the outside world without your approval. When such work reaches an outward step, like sending an e-mail, the step is not taken; the prepared call lands as a waiting proposal in the Recent notes list on Orbi's page. When an admin says Approve and send, the call goes out exactly as it was prepared; a proposal that uses a member's personal connection can be sent only by that member themselves. An unwanted proposal is closed with Don't suggest again, and a proposal that gets no response for two days lapses on its own. Research and reading stay open — the boundary applies only to steps that write to the outside. If you re-queue one of these cards yourself, the work is now running on your decision and the boundary no longer applies.

#When Orbi opens a chat with you

When a matter can be settled by one person alone, Orbi may open a private chat with them: to raise your long-waiting work one-to-one instead of in front of everyone, to settle a decision only you can make, or to discuss a sensitive correction in private. When the chat opens you get an "Orbi opened a chat with you" notification that takes you straight to the conversation; one notification per chat, no more. The opening message states the matter concretely and says that once it is settled, Orbi will carry the conclusion back to the team's shared surface — a private chat is a side path to a decision, and knowledge does not stay buried there. The team's shared agenda never moves into a private chat. At most one chat per person per day. If you do not want this, turn off the "Orbi may open chats with me" switch on Orbi's page and Orbi will never open a chat with you again.

You can look at the movements Orbi reads yourself, in the same place it does. The Activity page in the left menu lists card movements, comments, record writes and runs newest first, which is the raw material the briefing rests on. Channel conversations do not enter that list; they stay in the channel itself. More in Activity.

To turn it on, open Orbi's agent page (Agents → Orbi), pick the channel and set the time and the level. Orbi sets up the rest itself and you will not see anything extra in the flow list. When Orbi checks in and how its last runs went are shown in the rhythm rows on the same page ("Run now" lets you try it without waiting for the time). The Recent notes list on that page shows what Orbi wrote on its own and whether it got a response. If you say Don't suggest again on a suggestion you did not want, Orbi writes that to its memory and does not bring it up again. The daily cap also limits how many messages Orbi may write on its own in one day.

Proactive mode can be turned on from the Team plan upwards. Details: Plans and security.

#Behaviors and housekeeping findings

Everything Orbi watches for on its own is a behavior. The Behaviors list on Orbi's page shows these behaviors, how far each one may go and how it was received in the last 30 days. There are three steps: Acts and reports, Proposes, you apply and Only tells. A behavior's step is fixed by the system; Orbi cannot raise it on its own. Removing a member, revoking a token or changing a setting is on no step: those are always yours to do. The morning briefing, the nudges and the weekly plan are in this list too; you switch them with their own settings above.

Housekeeping behaviors look at the workspace itself and use no model, so they spend no credits. Every morning, once after 07:00 in the workspace's time zone, Orbi checks and writes what it finds to the From Orbi tab of the Inbox. Only admins see these findings:

• members who have not signed in for 30 days,
• tokens unused for 60 days,
• a day that used at least three times the last two weeks' daily average of credits,
• flows whose next run will fail because the person who set them up has left,
• storage that is 80% or 95% full.

There is one more: cards nobody has touched, a backlog or to-do card nobody has edited or commented on for 45 days. This proposal goes to the card's person, not the admins: it lands in the From Orbi tab of its assignee, or else its owner or the person who opened it, as "shall we close it?". The same question shows on the card itself: open the card and an "Orbi asks about this card" box sits at the top, where you can decide too. Close moves the card to done, and Undo puts it back where it was. If someone edited the card in the meantime, Orbi does not close it.

Findings and proposals meant for the admins are answered by one admin: Orbi's contact, set with Who should Orbi ask? in the proactive settings; if nobody is picked, it is the admin who turned proactive mode on. The other admins see the row, but only the contact has the buttons. The row says who decides and, once decided, who did what ("Ayşe applied it"). Questions wait for the contact and never move to someone else on their own. If the contact has not signed in for 14 days, Orbi tells the other admins so they can pick someone else. When there is something new, the contact gets a single "Housekeeping note from Orbi" notification. Orbi changes nothing itself in these findings. When you have fixed the problem, choose Done; if the problem goes away on its own (the member signed in again, the token was revoked) Orbi closes the finding itself the next morning. A finding you mark Don't suggest again does not come back. You can turn each housekeeping behavior off with its switch in the list. The Behaviors list is being turned on for workspaces gradually; if you do not see it yet, it is not on for your workspace yet.

#Orbi's night notebook

If the agent computer is also on in your workspace, Orbi works at night on its own computer. Between 02:00 and 06:00 in the workspace's time zone it reads the last two weeks of cards and comments in Genel and the open spaces, the last week of channel talk and the file list. It writes nothing at night: it works with read-only tools and uses nobody's personal connection. Closed spaces stay out of sight at night too. What it finds, it prepares for the morning:

• Following up on promises. A promise made in a comment or a channel whose time has passed ("ready by Friday"). In the morning Orbi reminds that person on the card, in their own words.

• Drift from a decision. A written decision (a meeting note, an agreement in a channel, a card) and work on the board going the other way. Orbi says it in the channel, quoting both exactly.

• Tidy files. Files that sit outside the folder their siblings live in. In the morning Orbi prepares a move for each and puts it in the admins' From Orbi tab. When you choose Apply the file moves right then; if the file or the folder changed since, nothing moves. Once you have applied five of its proposals in a row, Orbi makes the next move itself and leaves an Undo button on the row. Undo once and it goes back to only proposing. The Behaviors list shows where Orbi stands on this work right now.

• Channel filter. At night Orbi also reads the channel talk:

  • A promise made in a channel whose date has passed: Orbi asks the person who made it in a private chat.
  • A question nobody answered for two days: Orbi passes it to the person who can answer it.
  • Work agreed in a channel with no card behind it: Orbi asks the person who took it on in their inbox, "shall I make it a card?". Choose Open the card and the card opens assigned to you.

  Orbi never answers the questions itself. If you turned off Orbi opening chats with you, it asks in the channel in one line instead.

• Looking ahead. At night Orbi also thinks beyond the jobs in the list. It may notice a holiday post nobody started, a due date that falls on a day off, or something the team does every month but has not done this month. It leaves these in the inbox as suggestions, at most two a night. When an agent's work is needed ("shall I have the Content agent draft the holiday post?"), choose Open the card and the card opens for that agent and the work starts. Nothing runs before you say yes.

• Calendar and where the team is. Orbi reads public holidays from a holiday calendar. It works out which countries matter from where the team works: your browser's time zone is saved on your profile, so there is nothing to enter. On a public holiday in the workspace's country, Orbi writes no briefing and nudges nobody, on every day of a several-day holiday; the weekly plan still comes.

• Your rules. In the Things to keep an eye on section of Orbi's page, write in one sentence what Orbi should watch inside the workspace. Example: "tell me when a card promised to a customer passes its date". Orbi looks every night and leaves what it finds in the author's From Orbi tab. It does not repeat the same thing within a week. Only admins write rules, and at most 10 can be active. If you want work done at a set time, set up a flow, not a rule.

• Preparing work for a card. When a card has a meeting coming up, a due date or a question nobody answered, Orbi notices it at night and picks the agent of yours that fits the work. In the morning it has that agent do the work without opening a card, such as research before the meeting or a first draft. It leaves the result on the card in a comment addressed to the card's person. At most two such jobs are picked per night and one is done per morning round, once a week per card. The agent only reads in this work. The behavior starts off, because each job spends the workspace's credits; turn it on from the behavior list on Orbi's page.

• Watching the outside world. Topics your team cares about, such as grant calls, a competitor's announcements or a change in regulation. Orbi starts watching nothing on its own. When it sees interest in cards or chats (for example TÜBİTAK applications on the board) it asks in the admins' From Orbi tab, "shall I watch this?"; choose Watch to start. You can also add a topic yourself: tell Orbi in chat "from now on, follow R&D grant calls open to small companies", or write the same sentence in the Watch topics section of Orbi's page. Only admins can start or stop a topic. Orbi proposes where to look: after the first check it shows the pages in the From Orbi tab and saves them on the topic when you choose Use these pages. If it is heading the wrong way, stop the topic and write it again in a clearer sentence. Each topic is checked once a week; when something new comes up it lands in the From Orbi tab, with a link when it is on one of the topic's pages and otherwise with its address to check before opening.

The morning briefing and the nudge round deliver promises, drifts from a decision, unanswered questions and prepared work. Watch topics, rules, looking ahead and talk-to-card go straight to the From Orbi tab; rows the night leaves, and their notifications, come no earlier than 07:00. One person gets at most five night rows a day. The same limits apply there: the daily cap, one write per card per day, and quotes checked against their source. Every night Orbi keeps a notebook: what it noticed, which promises and decisions it is watching, and what it has already said. Anyone can read the notebook in the Night notebook section of Orbi's page. The night run spends credits; an admin sets a monthly budget in the same section (500 credits by default). When the budget is spent there is no night run that month; housekeeping findings go on, since they spend no credits. The night delivers to proactive mode's channel; if that channel is in a closed space, the night does not run.

#Assigning a specific agent instead of the orchestrator

Not every task has to go to the orchestrator. If you know exactly what you want, assign the task straight to a specific agent. The planning layer is skipped and the work finishes faster and with fewer tokens.

Keep the orchestrator for work that needs more than one specialism. Details: Tasks and Kanban.