Agents
Agents that prepare work for one of your roles, within a budget, and wait for your approval.
Last verified 2026-10-06
In the web app
What an agent is
An agent works for one of your roles on a schedule and prepares one kind of result, which waits on your approvals desk.
Agents only prepare. You approve every result: an agent never sends, publishes, buys or changes anything itself.
Every agent works at the level Prepare, and no other level is offered.
Open tasks check
The first agent, Open tasks check, looks at a role's open and overdue tasks every Monday at 07:30, Zurich time.
When tasks are overdue it proposes at most one task, such as "Review 4 overdue tasks in Finance", with a reason and a due date.
Approve on the approvals desk adds the task to the role; Reject drops it, and nothing is created.
Tender watch
The Tender watch reads new public tenders on SIMAP for a role and proposes the ones that fit its watch list; see Tender watch.
Content creator
The Content creator writes drafts in your voice from what you select, and only when you ask; see Content creator.
LinkedIn drafts plans a few LinkedIn posts a week from what you approved, for you to copy and post yourself; zonask never posts. See LinkedIn drafts.
SEO and GEO agent
The SEO and GEO agent reads a role's Search Console numbers every Monday, asks a small panel of AI prompts whether they cite your sites, and proposes at most 5 findings; see SEO and GEO agent.
Turning agents on and off
Under Settings, Agents you see each of your roles with its agents, whether each is off, on, paused or has reached its budget, and its last runs.
Turn on starts an agent for a role; Run now starts a run within a minute; Pause and Resume stop and restart one agent, and a role has its own pause for all its agents.
Pause all agents stops every agent of yours before its next step, and the page then says "All agents are paused."
A run that was stopped does not continue when you resume; the next run starts fresh.
Budgets
Each agent has a monthly budget of 5 Swiss francs, and all your agents together 20 Swiss francs a month; a run uses at most 200'000 tokens.
Before every call to the AI an agent sets the most it can cost aside, and afterwards books what it really cost, so a budget is never passed unnoticed.
When a budget is reached the run stops, the agent pauses until the next month, and Today tells you once.
Monthly limit
Each agent's own monthly limit can be raised or lowered, from 0 up to the limit every agent shares (20 Swiss francs by default); 0 means the agent may not spend.
Set it on the Agents tab of Runs, under each agent; the tender watch's own form has the same field.
Raising it restarts an agent paused for its budget right away, instead of waiting for the next month.
What a run log keeps
For each run the page shows when it ran, how many steps and tokens it used, what it cost and how it ended, for example "Stopped: budget reached".
A run log keeps ids, counts, times, costs and fixed outcome codes; it never keeps what a tool read or what the AI wrote.
An agent's working notes and a proposed task are encrypted and deleted after 30 days; an approved task stays as a task.
The operator of zonask can stop every agent of every member at once.
Delete all my data removes your agents, their runs and their proposals.
Results and cost per approved result
The Agents tab of Runs shows each agent's last 8 weeks, what it cost per approved result and this month's spend per role; see Roles on Headquarters.
When you reject an agent proposal, Why did you reject it? (optional) offers Not relevant, Wrong facts, Wrong tone and Duplicate. Runs shows the rejection counts by reason. Only the chosen code is kept in the audit entry.
An agent quality alert appears if its approval-without-edit rate falls by more than 20 percentage points between the last two completed weeks and the two before, with at least five proposals in each period. A cost alert appears if cost per approved result at least doubles, with at least three approved results in each period. Each alert appears once per agent and week, links to the agent and appears in weekly tuning until cleared.
The run room
Runs has "Running now and next": every agent that is on, with its own schedule in words, when it next runs and one status, Running now, Waiting for its next run, Paused, or Paused: this month's budget is used up. Recent runs lists your agent and standing-order runs together, newest first, with the cost of each agent run and, when one failed, one sentence why. Run again offers a failed agent run only while it is that agent's newest run and the agent is neither paused nor out of budget; it starts the same way Run now does, so both refuse a paused or budget-stopped agent with the same sentence. See Roles on Headquarters for the Agents tab.
In Claude or ChatGPT
list_agents shows your agents per role and get_agent_runs their last runs, as a list where the host shows cards. Each agent now also carries its schedule in words, its next run and the same status the run room shows.
run_agent_now, pause_agent and resume_agent start, pause or resume an agent after your confirm; run_agent_now refuses a paused or budget-stopped agent with the same sentence the run room's Run again shows.
update_agent with month_cap_chf alone raises or lowers one agent's own monthly limit, 0 to the limit every agent shares; it shows the amount first and saves it only after your confirm. The amount is in francs and cents: one with more than two decimals is refused.
list_runs also lists your recent agent runs beside your standing orders', each with its cost and, when one failed, why; ask "What did my agents do this week and what did it cost?"
What an agent proposes waits for you in list_approvals, and answer_approval approves or rejects it with an optional reason code. get_agent_stats shows its rejections by reason.