Skip to content

Constraint: every fact, on one page

This page holds every fact on constraint.com in plain words, for people who want the detail and for AI assistants and search engines. It follows the site's story in the site's order. The site's screens point to its sections by short links such as everything#pilot; each section's link is its heading in lower case with hyphens.

Steer by intent, at scale, together. That is Constraint's line, and the order of this page.

AI at work is still single-player: one person, one thread, one chat at a time, with work no one else can see or build on. Constraint is built to change that.

All 20 sections
  1. What it is
  2. Who it's for
  3. Steer by intent
  4. At scale
  5. Desktop agents
  6. Connectors
  7. Together
  8. Assign a question
  9. Share with other teams
  10. Analysis Hub
  11. Your AI account
  12. Where data goes
  13. IT questions
  14. Pilot
  15. Price
  16. After the pilot
  17. Who builds it
  18. Glossary
  19. FAQ
  20. How to start

What it is

Constraint is a desktop app a whole team shares. Each person runs it on their own computer. It is a department of AI agents, directed by the whole team.

You tell one AI, the VP, what you want, the way you'd brief a team lead. The VP turns that into tickets in the ticketing service your team already uses and puts an agent team on each ticket. The agent teams work in parallel.

Home, each person's own page in the app, shows them only the decisions that are theirs, as question cards with the research behind them. The AI waits for the answer, then carries on and notes it on the ticket.

The work happens in the open. Teammates see each other's AI sessions live, with whose computer runs each, and can step in. Setups, meaning departments and sentries (automatic checks that find work), are shared across the team. Every session's record, called a trace, is saved with its ticket, and everything the team has done makes every run better.

The three pillars of this site, and of the product, are Steer by intent, At scale and Together. Each has its own page. This page holds everything under each, in that order.

Who it's for

Constraint is for teams that run on tickets and already use AI seriously, and who want to steer it by intent rather than babysit it: say what you want once, see the work filed as tickets with an agent team on each, and answer only the decisions that are yours.

It is also for a team lead who has never used AI. You brief the VP in plain words, as you'd brief a colleague, and it asks you only the questions that are yours to answer.

When Constraint fits. Your team's work lives on tickets. Your people work on their own computers, with files, tools and accounts there that the work needs. Several people need to see, join and build on each other's AI work. You want to start on real work with the people who build the product; the price is agreed on a short call.

When something else fits better today. You want to sign up on your own and start the same day. You want agents that keep working with every computer off. You want a library of ready-made templates, or spend caps and usage dashboards. Constraint does not offer these, and this site does not claim them.

Steer by intent

Working with AI usually means sitting in a thread, in the details, one chat at a time. With Constraint you work at the level of intent: the VP turns what you say into tickets, sentries find the work you didn't ask for, and every question comes back to you in one place.

  • Say what you want. The VP is the one AI you talk to. It runs your AI department. You brief it in plain words; it aligns with you on what you want, files the tickets in your ticketing service, and puts an agent team on each. The agent teams work in parallel. If no role fits the work, the VP asks you before it creates one.
  • The work stays with its ticket. Each ticket holds its AI sessions, so you can follow the work without keeping track of separate chats. Every session's record is saved with its ticket.
  • Every question, one place. Every question for you comes back in one place, on Home, whether it comes from aligning with the VP or from a worker mid-task, so you never jump between threads. Home shows each person only the decisions that are theirs, as question cards with the research behind them. The AI waits for the answer, then carries on and notes it on the ticket.
  • Earlier answers first. Agents check the team's earlier answers before they ask anyone.
  • Sentries find the work. Automatic checks, called sentries, do discovery and analysis on their own schedule. They watch what you point them at, find the next piece of work, and file or propose tickets for the department's agent teams. A proposed ticket waits for your yes on Home.

At scale

Tickets now arrive as fast as you can think. Every one done well. So every ticket gets an agent team, every agent is live on its ticket, the agents run where your work is, and they have the skills and connectors to do the job right.

  • Agents, for every ticket. The VP puts an agent team on each ticket, and the teams work in parallel: hundreds of tickets, each with its own agent team. Every agent shows live on its ticket, on any teammate's computer, with whose computer runs it.
  • Built like your team. A department is one part of your company, like marketing, with its own AI roles. A role is a job, with its how-tos, called skills, and its AI model, chosen from your own AI account. An agent team runs roles in order, for example a planner, a writer and a reviewer. A ticket goes to one role or to an agent team.
  • Equipped to do it right. Each department has rules that every role reads first. Each person chooses how far agents go before they ask: ask when permission is needed, allow edits, or plan without editing. A permission can be allowed once, or for the rest of a session. A role can be given checks to run before it says it's done. You choose these settings. The AI stops to ask rather than guessing. Connectors give agents your team's tools, and the department's skills give them its know-how, so they can do the job right without coming back to you. Skills are saved in versions; when someone improves one, every role that uses it can take the new version. See Connectors below.
  • Tickets that start themselves. Auto mode: assign a ticket to yourself, and the right agent team starts on it by itself, in the stages you allow. You choose which stages auto mode starts work in. Work already running carries on if its ticket moves. It asks on Home when it needs you.
  • Desktop agents, not cloud agents. Each agent works on a teammate's computer, with the files, tools and accounts there. See Desktop agents below.
  • Agents work one ticket together. Agents on different teammates' computers work one ticket together.

Desktop agents

Constraint's agents are desktop agents, not cloud agents. Each runs on a teammate's computer, with the files, tools and accounts there as well as the team's online tools. Agents run where your work is.

That is why they're stronger. They work with what's on that computer, which a cloud agent can't reach: the files, the tools and the signed-in accounts. And the whole team can see every one of them working, live, on its ticket, with whose computer runs it.

Agents can use the AI subscription you already pay for, instead of paying for each use.

Each person's agents work on their own computer while the desktop app is running. Each session runs on the computer of the person who started it. A teammate can join a session only while both computers are on. Agents do not work while a computer is off; see the FAQ.

What is saved and shared: every session's record, its trace, goes to Constraint's cloud service with its ticket, so the whole team's AI can build on it. See Where data goes.

Connectors

Connectors are the tools agents connect to. Constraint connects to over 400 tools today, through an integration provider. An administrator approves each connector, and each person connects their own accounts.

Connectors give agents your team's tools, and the department's skills give them its know-how, so questions to you stay rare. Your team's ticketing service is one of the connectors, and it is connected in the first week of a pilot.

Does it work with your tools? Type a tool's name to see whether it's one of the over 400. The finder only searches the list: your typing stays in your browser, and nothing is connected from here. List updated 29 October 2026.

Together

With this much work commissioned and running, two things become true: not every question belongs to the person who started the work, and no work is perfect the first time. So each question goes to the person it belongs to, teammates step into each other's sessions, teams share what works, every agent builds on the team's work, and every run teaches the next.

  • Ask the right person. When your AI asks something a teammate should answer, assign it to them. See Assign a question below.

Step into the work. Ask a teammate, step into a live session, or share one ticket, each from your own computer.

  • Ask a teammate. Your AI's question goes to the person who should answer, and the answer comes back to your session. See Assign a question below.
  • Step into a live session. Teammates see each other's AI sessions live, with whose computer runs each. They can message a session, answer its questions, stop it, or step in, while both computers are on. Each session runs on the computer of the person who started it.
  • Several people on one ticket. Several people can be on one ticket, and the AI working it reads their comments. Agents on different teammates' computers work one ticket together. Owners, comments, mentions and dependencies keep people around each ticket, and the agents working it read the comments.

Accelerate together. Every run is remembered, looked up, shared and learned from, so the whole team gets faster.

  • Every run, remembered. Every session's record, its trace, is saved with its ticket, so every teammate's AI can build on the team's work. A Traces page shows them all. To hand a session over for good, export its trace; a teammate imports it on their own computer and carries on.
  • Looked up before they ask. Agents look up how and why past work was done before they ask you anything. They look up earlier tickets and sessions within the team access an admin sets, so they can keep to earlier choices that were deliberate and ask only about ones that look accidental.
  • Share what works. Within a team, its departments and sentries are shared. An admin can share them with other teams too. See Share with other teams below.
  • Learn from your best work. The Analysis Hub reads the whole team's traces, finds what isn't working, and proposes a better version. See Analysis Hub below. An admin can also review the traces of one kind of work and raise the baseline from the best runs.
  • Who can change what. Only the owner changes a shared role, sentry or department; everyone else makes a copy.

Assign a question

When your AI asks a question that a teammate should answer, you can assign it to them. They get it in their app and answer on their own computer, even from home, and the answer reaches the session on your computer, which carries on. A decision waits for the person it belongs to, not whoever's computer runs the work.

A teammate can answer a session's questions from their own computer, but approving a tool is never done from another computer.

One person can also answer a question on their own app without being asked, and the answer lands in the teammate's session, which carries on with it.

Share with other teams

Within a team, its departments and sentries are shared. An admin can also choose to share a department's setup and its sentries with other teams. Only the owner changes the original; everyone else makes a copy. Skills are saved in versions, so when someone improves a shared how-to, every role that uses it can take the new version.

Analysis Hub

The Analysis Hub reads the traces of a person or a team, shows which skills, roles and agent teams aren't working, and proposes a better version. Every teammate's sessions are saved, so the Analysis Hub reads the whole team's runs, not one person's.

Alongside it, an admin reviews the records of one kind of work and raises the baseline from the best runs. Where people keep stepping in becomes the next version of a skill, and requests that keep coming back become a sentry or a new role.

Your AI account

Connect your own AI account and choose from its models. Each session runs using your own AI account. Each role in a department has its own model, chosen from that account. Agents can use the AI subscription you already pay for, instead of paying for each use.

Where data goes

  • Each session runs on the computer of the person who started it, using your own AI account.
  • Its record, the trace, is saved to Constraint's cloud service with its ticket, so every teammate's AI can build on it.
  • Before a session's record is saved to Constraint's cloud service, screenshots are removed and keys, tokens and passwords are blanked out.
  • People see session records only for the teams they can read.
  • Session conversations are stored as traces on your computer and in Constraint's cloud service.
  • Session records in Constraint's cloud are deleted 30 days after the account is deleted. Today, Constraint uses them only to run the service.
  • Your AI service and connected tools receive what each session sends them.
  • Tickets live in your team's ticketing service, where the VP files them.

IT questions

Five plain answers, and a few more, from the facts above.

  • Whose AI account does it use? Your own. You connect your own AI account and choose from its models.
  • Who approves what gets connected? An administrator approves each connector, and each person connects their own accounts.
  • What leaves the computer? A session's record, its trace, saved to Constraint's cloud service with its ticket. Before it is saved, screenshots are removed and keys, tokens and passwords are blanked out. Your AI service and connected tools receive what each session sends them.
  • Can a teammate approve a tool on my computer? No. A teammate can answer a session's questions from their own computer, but approving a tool is never done from another computer.
  • Who sees the records? People see session records only for the teams they can read.
  • How long are session records kept, and what are they used for? Session records in Constraint's cloud are deleted 30 days after the account is deleted. Today, Constraint uses them only to run the service.
  • Who can change a shared role or sentry? Only the owner. Everyone else makes a copy.
  • How far do agents go before asking? Each person chooses: ask when permission is needed, allow edits, or plan without editing. A permission can be allowed once, or for the rest of a session.
  • Where do agents run? On each person's own computer, while the desktop app is running. They do not work while a computer is off.

Pilot

Every engagement starts with a pilot on one team's real work, under a short evaluation agreement. The pilot's length is agreed with you on the short call, and it starts with the first two weeks. You work directly with the people who build Constraint. On your side, that's the team and the person who owns its work.

The first two weeks:

  • Day one. We agree with you on the first two pieces of work to hand over, and how we'll both know they're done.
  • Week one. We connect your team's ticketing service, turn your docs into your department's roles and skills, and set up each person's computer.
  • Week two. The first work runs on your real tickets, and we train the team hands-on, leaving it resources to work from.

If your first work isn't running on your real tickets by the end of week two, you can stop and you won't pay for the pilot.

The weekly hour. An hour-long meeting with us every week, during the pilot and after it.

Every pilot includes the setup above and the weekly hour.

Price

Every pilot includes the setup and an hour with us every week. We agree the price with you on a short call, and it follows the work as your needs change.

After the pilot

When the pilot ends, we review it together. If it isn't the right fit, you can stop there. If you continue, ongoing work runs under a services agreement, and more teams can join. Each session runs on your own AI account. The hour-long meeting with us every week carries on after the pilot.

Who builds it

Constraint is built by its founding team. During the pilot and after it, you work directly with them. We set up every pilot ourselves, so we start a few teams a month.

Glossary

The app uses short names for a few things. Here is each one in plain words.

  • The VP. The one AI you talk to. It runs your AI department: you brief it, it aligns with you, files the tickets and puts an agent team on each.
  • Home. Each person's own page in the app. It shows only the decisions that are theirs, as question cards with the research behind them. When nothing needs you, it says so.
  • Department. One part of your company, like marketing, with its own AI roles, agent teams and the rules every role reads first.
  • Role. One kind of AI worker, like a planner, a writer or a reviewer. A role is a job, with its skills and its AI model. A ticket goes to one role or to an agent team.
  • Agent team. A few roles that work a ticket in order.
  • Skill. The written how-to a role follows. Skills are saved in versions; when someone improves one, every role that uses it can take the new version. Each role uses the version you choose.
  • Sentry. An automatic check that does discovery and analysis on its own schedule or when something happens in a connected app, finds the next piece of work, and files or proposes its ticket.
  • Trace. The saved record of one session: everything the AI did on a ticket. Traces are saved with their tickets, and a Traces page shows them all.
  • The Analysis Hub. Reads the traces of a person or a team, shows which skills, roles and agent teams aren't working, and proposes a better version.
  • Ticket. One piece of work in your team's ticketing service. The VP and sentries file tickets there.
  • Session. One stretch of AI work on a ticket. It runs on the computer of the person who started it, and the whole team can see it live.
  • Agent. One AI doing a job in a session, on a teammate's computer. A desktop agent, not a cloud agent.
  • Connector. A tool agents connect to. An administrator approves each connector, and each person connects their own accounts.

FAQ

What is Constraint? A desktop app your whole team shares: a department of AI agents, directed by the whole team. You tell one AI, the VP, what you want. It files tickets in your ticketing service and puts an agent team on each, and each person answers only the decisions that are theirs.

What's a department? One part of your company, like Marketing, with its own roles and agent teams, directed by the people on that team.

What's a role? An AI worker with its own job, skills and model. A ticket goes to one role, or to an agent team that runs roles in sequence.

What's a skill? Written instructions a role follows. Skills are saved in versions, and each role uses the version you choose.

What's a sentry? A check that spots recurring work and proposes or files a ticket for it. It runs on a schedule, or when something happens in a connected app.

Do I need to know about AI to use it? No. You brief the VP in plain words, the way you'd brief a colleague, and it asks you only the questions that are yours to answer.

Where do the tickets go? Into the ticketing service your team already uses, so everyone sees the plan. Every session's record is saved with its ticket.

Do agents keep working when my computer is off? No. Each person's agents work on their own computer while the desktop app is running. A teammate can join a session only while both computers are on. Work can be handed over for good by exporting a session's trace, which a teammate imports and carries on.

Can my teammates see what my AI is doing? Yes. Teammates see each other's AI sessions live, with whose computer runs each. They can message a session, answer its questions, stop it, or step in while both computers are on. You choose who can step in, per session and per person; each request to join is checked first.

What if my AI asks something a teammate should answer? Assign the question to them. They answer in their own app, on their own computer, and the answer reaches your session, which carries on.

Can several people work on one ticket? Yes. Several people can be on one ticket, and the AI working it reads their comments. Agents on different teammates' computers can work one ticket together.

Who can change a shared role, sentry or department? Only the owner. Everyone else makes a copy. An admin can choose to share a department's setup and its sentries with other teams.

Whose AI account does it use? Your own. You connect your own AI account and choose from its models.

Where does our data go? Each session runs on the computer of the person who started it, using your own AI account. Its record is saved to Constraint's cloud service with its ticket, with screenshots removed and keys, tokens and passwords blanked out. Session conversations are stored as traces on your computer and in Constraint's cloud service. Session records in Constraint's cloud are deleted 30 days after the account is deleted. Today, Constraint uses them only to run the service. Your AI service and connected tools receive what each session sends them.

Does it work with our tools? Constraint connects to over 400 tools today, through an integration provider. An administrator approves each connector, and each person connects their own accounts.

What does it cost? Every pilot includes the setup and an hour with us every week. The price is agreed on a short call.

How long is the pilot? It's agreed with you on the short call. It starts with the first two weeks: day one, week one and week two.

What happens in the first two weeks? Day one: we agree the first two pieces of work to hand over, and how we'll both know they're done. Week one: we connect your ticketing service, turn your docs into roles and skills, and set up each person's computer. Week two: the first work runs on your real tickets, with hands-on training.

What happens after the pilot? We review it together. If it isn't the right fit, you can stop there. If you continue, ongoing work runs under a services agreement, and more teams can join.

Where are the customer stories? There are none on this site. What you can check instead: the real app, shown on the site; a sample task's full history, from request to finished work; the exact rules on who can change what, where data goes and when agents work; the pilot, with a day-by-day plan for the first two weeks; and the people who build it (Who builds it).

How to start

Book a short call and tell us about your team and the first work you'd hand over. The call is 30 minutes, with the founding team. On the call we agree the price. Then the pilot begins: on day one we agree the first two pieces of work, in week one we set up your team, and in week two the first work runs on your real tickets. See the pilot page.

Four questions for a useful hand-off help choose the first two pieces of work: goal, information, review and hand-off. When you book, please leave out confidential business information, company documents and AI sessions; a short description is enough.