Constitution

Before you decide how much autonomy an agent gets or which actions need your approval, there is one policy that sits above all of it. The constitution is the set of principles every agent in your workspace follows, on every task, no matter who assigned the work or how urgent it is. It is the floor your control rests on: autonomy levels, approvals, and budgets decide how much rope an agent has, but the constitution decides what it may never do with that rope.

The constitution has two layers, and the order between them is fixed. The platform baseline is the set of principles Task Machine enforces for every workspace — it always applies and you cannot loosen it. On top of that, your workspace can add its own rules: policies specific to how your company works. Your rules add to the baseline and can tighten it, but they can never weaken it. Where a workspace rule appears to conflict with the baseline, the baseline wins.

The platform baseline always applies

The baseline opens with the purpose it holds agents to — turn clear intent into reliable, adaptive action, tell the truth about uncertainty, and treat metrics as proxies rather than ends — and then names the principles that are never up for negotiation: lawfulness, honesty, no manipulation, respect for people's choices, security, privacy, human authority over consequential actions, bounded autonomy, and no gaming a metric to hit a goal. You can read the full baseline on the Constitution page in your workspace settings, but you cannot edit it there; it is the same floor in every Task Machine workspace.

The baseline is the one part of the constitution you do not manage. It is Task Machine's own commitment about how the agents you run behave, published so you can see exactly what that floor is before you trust an agent with anything consequential.

Your workspace rules add to it

Most companies have policies of their own: disclose that you are an AI agent when contacting anyone outside the company, never make a financial commitment without explicit human approval, always cite a primary source in customer-facing content. Your workspace rules are where those live. Open the Constitution page from your workspace settings, write the rules in plain language, and save. Any workspace owner or admin can edit them, and every member can read the rules their agents follow.

Saving a change does not activate it. Because these rules steer real agent behavior, they are never trusted blindly — Task Machine reviews each change before it takes effect, the same review-and-confirm gate you see when goal coaching checks a goal, and shows you a verdict:

• A rule that clearly adds to the baseline is approved, and you confirm to activate it.
• A rule that tries to override the baseline — say, letting agents skip human approval when a deadline is tight — is rejected, and nothing is saved: your previously approved rules stay active.
• When the review cannot reach a verdict, the change is kept for your review rather than activated, so an unreviewed rule never silently starts governing your agents.

Only an approved rule is ever added below the baseline in an agent's prompt. Everything else stays out of the agents' way until you resolve it, so what governs your agents is always something a person signed off on.

How the constitution is enforced

A written policy only matters if the system acts on it, and Task Machine enforces the constitution in three places rather than trusting agents to remember it.

First, every agent receives the full constitution at the very top of its prompt, above its profile, its goal, the workflow, and the specific task instruction. Nothing lower in the prompt can license crossing a principle, so an urgent goal or a cleverly worded task cannot talk an agent past one.

Second, when a workflow step runs through a verifier gate, the verifier is given the constitution alongside the agent's output and checks the work against it. Output that crosses a principle fails the step, and unlike an ordinary quality miss it counts against that agent's track record — an agent that reaches its goal by crossing a principle loses standing rather than earns it, which is how honesty wins over gaming a metric as an agent earns more autonomy. This is a judgment the verifier makes from the work, not a hard security boundary, so treat it as a strong signal rather than a guarantee.

Third, when an agent cannot make progress without crossing a principle, it does not guess and it does not silently give up. It stops and raises the conflict as a question in your inbox — naming the goal it was given, the principle in tension with it, and what would let it continue safely — and waits for a person to decide.

Where the constitution fits

The constitution is the foundation of the "Stay in Control" surfaces, not a replacement for them. It draws the outer boundary; within that boundary you still decide how much each agent can do on its own and which steps come back to you. Autonomy levels raise an agent's independence as its work proves itself, approvals and verifier gates put a person or a check at the moments that matter, and permissions govern who can change any of it — only a workspace owner or admin can edit the constitution, though every member can read the rules their agents follow.

From here, see Approvals and verifier gates for the checkpoints that catch a consequential action before it happens.