Pause and Safety

An autonomous system is only as trustworthy as your ability to stop it, inspect it, and bound it. The platform takes all three seriously, and it starts from a conservative posture: every agent ships disabled, and Impulse, the autonomous layer, is off until you turn it on. Autonomy is something you grant, in layers, at your own pace.

The Pause Layers

Global pause is the big switch: one click in the sidebar, or "pause everything" said to the master agent, stops all autonomous activity at once. Under the hood it is a file, work/.global_pause, create it and the system holds, delete it and the system resumes, which means even a shell session can stop the platform if that is ever the closest tool.

Per-agent pause works the same way for one agent: the pause button on its lane, or a work/.pause_<id> file, holds that agent, Impulse included, while everything else continues. This is the everyday control, pause the marketing agent during a pricing change, pause social during a sensitive news day, without touching the rest of the system.

Tool-level pauses add a subtler option for outward-facing areas: php tools/marketing.php pause and its siblings stop sends without stopping the agent, so list processing and reply reading continue while nothing leaves the building. Config-level paused flags exist per area as well, and resume is honest about the stack: resuming surfaces whichever pause still applies, so a paused layer never hides behind a resumed one.

Kill Now is the immediate option: every agent lane carries it, and it stops the agent's current task on the spot, mid-run, while the agent itself stays registered and ticks again on schedule. Pauses are for plans, Kill Now is for right now.

Structural Guarantees

The outward-facing presets carry safety properties that hold regardless of what any model decides, because they are enforced by the system rather than requested of the AI. The social media preset defaults to review mode where every outgoing item queues for your approval, and replies to large accounts always queue, permanently. The customer service preset answers only from verified knowledge, no match, no answer. The marketing preset enforces unsubscribe compliance on every email and checks suppression before every send. These are properties of the pipes, not promises of the agents.

One more boundary is written into every agent's instructions: agents may improve their own environment, including adding new tools as they learn what they need, but they are never to stop or break the running system itself. Self-extension is welcome, self-interference is out of bounds.

Flags: The Human-Attention Mechanism

Flags are how the whole platform asks for you. When any agent meets something that needs human judgment, it writes a flag, a file in threads/flags/, carrying a type, a source, and a topic. Three types cover the territory: unanswered_question, something was asked that the system cannot answer from verified knowledge; human_judgment, a decision is yours to make, a legal-sounding reply, a high-intent customer, an ambiguous situation; and system_observation, something you should know about, noticed by an agent or by Impulse on its rounds.

The flag counter lives in the sidebar on every screen, and the flags panel shows each item with its context. You can also handle flags conversationally, the master agent routes them, takes your answers, and files the results, and answers to unanswered questions flow into the knowledge base so each one is settled permanently. Flags are the inverse of alert fatigue: agents handle what they can verify and escalate what they cannot, so the queue stays short and everything in it genuinely deserves you.

The Trust Gradient

The layers compose into a deliberate on-ramp. Day one: agents enabled one at a time, social in review mode, Impulse off, you watching the activity feed. As the work proves itself, you extend the leash, longer ticks become shorter, review queues relax where the edit rate has fallen to zero, Impulse comes on. At every stage, the global pause is one click away and Kill Now is beside every lane, so the cost of an experiment in autonomy is never more than the moment it takes to stop it. Trust is granted in increments and revocable in one.