Questions | B Team (working title)

We keep building things nobody uses. How is this different?

Every piece of work is meant to start from your goal and a done-when, and to end with real users and a "used after release" number. That is the design. None of it is built. Planned, not built Today the triage of new work is measured, and nothing more. Build what actually matters

Am I locked into one model vendor?

Which model a role uses is a setting, not code, behind one gateway. Measured Three providers answer, two are wired to the gateway, and all four decided roles use Anthropic today. The reviewer is not yet a different model family from the worker (issue #82). Planned: a model is allowed only if your data policy allows it and it passed the role's test set. Planned, not built The Scaleway key answers (16 models) but is not connected to the engine, and it is replaced before it is (issue #81). Planned, not built Choose your models, keep your freedom

I have no engineer. Who reads the code?

Today there is no step where an engineer must read it. A separate reviewer looks at the work, and the person who accepts it decides. Measured (pair runs) The human review level lets you choose. At level A, checks only, every change carries the line "no human engineer has read this code". Level A is decided and not built. At level B, a named engineer reads risky changes. At level C, an engineer reads every merge request. Planned, not built The case that breaks the gate: you approve a change nobody on your side can judge. People and AI, building together

What stops an agent from merging or publishing on its own?

Gates are in the machinery. An agent cannot give itself the approval: forged accepts and edited comments are refused. Measured, with a recorded deviation Assistants cannot answer a gate. Built A worker holds no write key. Measured (in part) The deviation: whoever holds a project's key (Temporal token) to the workflow service can send a decision under any name (issue #32).

Does the reviewer actually catch mistakes?

Not measured yet. The pair runs and the reviewer cannot change the work. Nobody has shown that a reviewer catches a real fault (issue #28). A reviewer from the same model family may share the worker's blind spots.

Where does my data go?

Prompts go to Anthropic and sit in plain text in the workflow service's history (issue #57). No data policy has been checked against any provider's terms. Keys sit in plain git-ignored files, and traces hold full words by default (issues #51 and #84). A European provider is the plan, in the order set out on the models page, and no role runs on one yet. We give this answer as it is. Use AI and stay compliant

My test users have no code host account.

Feedback and user tests are planned so they need none. Planned, not built Today the mailbox counts an item only from a member of the project, by design, so a stranger's text never starts work. A test user's answers would be written in by the engine and marked as outside text.

Isn't this more agents to babysit?

It is a fair worry. The answer is a separate reviewer and one inbox for everything that needs a person. The inbox is one per project today, not one across projects. Nobody has measured the time a person spends reviewing. That is the design goal, and it is not proven.

Do I need to code?

You must not need to, and that is the goal. Today you need a terminal to install and run it. The web page shows work and sets up a project from a form, but it cannot start or stop anything. Built Accepting a piece of work is a comment in GitLab today.

Do I need GitLab?

Yes, today. The inbox and the mailbox are GitLab issues. Self-managed GitLab as a first host, and other hosts behind one layer, are planned (issue #279). Planned, not built

What will it cost me, and can it run away?

A spending limit is refused before a call is made. Measured One switch stops all spending, but it does not stop a job already running. Limits on steps, tokens and wall time for each job are planned (issue #58). One provider key and one bill serve all projects today. Anthropic's pricing page says batch saves 50 percent (tool summary, 2026-10-13); nothing here measures it on our jobs. Fit your people and your budget

Can I run it myself, and is it hosted in Europe?

You can run it on a MacBook with Docker Compose. Built A guided setup into your own account or server is meant to be the first offering, and it is not built. There is no guide beyond the project's own readme. It is not hosted in Europe. We intend to offer a hosted service in Europe, after its preconditions are measured. Planned, not built

Is it finished?

No. It is an early build. Two job types, batch work and the mailbox, run against test setups. It is not offered to anyone yet, and "B Team" is a working title. What is not built yet

Does B Team make us compliant with the GDPR and the AI Act?

No. The duty stays with your organisation, and this is not legal advice. B Team is meant to produce the records your compliance needs as you work: who asked for what and who approved it, which AI produced each output, who checked it, which providers may see which data, and a record nobody can quietly change. The base is measured. The rest is not built. Planned, not built Use AI and stay compliant

Do we need a processing agreement with you?

It depends on whether B Team can reach your personal data. If you run it in your own account and B Team has no access, it is probably not a processor. That is the research report's reading of the European Data Protection Board's examples, and a lawyer must confirm it. If B Team hosts, or has support access, it is a processor, and Article 28 of the GDPR needs a written contract. A template and a default of no access are planned. Planned, not built

We hold ISO certificates. Will B Team get us through the audit?

No. B Team is meant to support your controls with records and a supplier package. Planned, not built It does not make you certified. B Team itself holds no certificate yet, and its assurance comes in stages.

How do I know an agent got it right?

You do not rely on the agent. No agent is trusted, and every output is checked. Everything an agent does is recorded and can be read back. Measured No change to an agent's instructions or model is used before it passes its test set. Today a person runs that gate. Measured; gate manual A separate reviewer checks, and a person approves. Nobody has shown that the reviewer catches a real fault (issue #28). Checks that run without AI, written by AI and approved by a person, are planned. Planned, not built How we raise reliability

Is it open source?

No, and this site never says so. The direction is source-available under a PolyForm licence. The working direction is PolyForm Noncommercial 1.0.0, and a lawyer confirms it (issue #22). The Open Source Definition says a licence "shall not restrict any party from selling or giving away the software" (tool summary, 2026-10-13), so a licence that bars resale is source-available. No rule the research read forbids buying a source-available tool. How it scores in tenders that score open licences first is a question for the lawyer.

Is it free for NGOs?

The software, yes, for verified non-profits. Decided Decided 2026-10-13; the licence is confirmed by a lawyer. You pay for onboarding, support for each active project, and yearly maintenance and support. Security fixes are free for everyone. The amounts are not decided. Not decided Fit your people and your budget

Who owns what the agents write?

Unsettled. The Dutch cabinet says AI output is not protected by itself, and no EU rule settles it for code. This site promises no ownership. A page on what the agents write for you, and what you may do with it, waits for a lawyer's answer.

Who do we call when something breaks?

There is no published contact yet, no vulnerability policy and no reporting desk. That is not built. Planned, not built Until then, a failure reaches your inbox, and the contact is the placeholder on the page of how to reach us.

Which languages?

English only today. Other languages are not decided.