Chief of Staff, and Agents You Build Yourself
Two shipped changes to Muin's AI: one short ranked list of what needs you today, and a wizard that builds an agent of your own with its own tunable settings.
Two things changed in how Muin’s AI reaches you. The first is that it now tells you what needs doing instead of waiting to be asked. The second is that you can build an agent of your own when nothing in our library does the job.
Chief of Staff
An assistant that only answers questions puts the work of knowing what to ask back on you. Every morning you open six dashboards to find out what changed, and the thing that mattered was on the seventh.
Sidebar → Intelligence → Chief of Staff is one short list instead. The few things that need you today, ranked, each with the action that clears it. Below it, everything else in order, with its deadline and its reason, so you can see what is coming without acting on it.
When nothing needs you, the page says so. It does not pad the list to look busy.
Coverage
You are about to travel. Who takes the approvals?
Coverage answers that in one screen: who is covering you, who it goes to if they do not act, the window, what is covered, and the amount above which a decision waits for you regardless of who is around. If nobody is covering you, it says Nobody is covering you, which is the answer you want before you leave, not after.
Readiness
An inspector is coming. Which of their questions can you not answer?
Readiness asks them: pick Tax authority, Fire marshal, Single audit or Insurer, and the page answers each question from your own records, with the evidence attached.
The part that matters is the part that refuses. A question with no record behind it is reported as a gap, never answered optimistically, and each gap becomes a task somebody owns. A readiness report that always says you are ready is worth nothing.
What if
Change one assumption and see what it does to the rest, against your real figures for the current fiscal year.
Every number cites its source. A lever the platform has no input for is badged Needs an answer first rather than filled in with something plausible. We would rather give you a partial model you can trust than a complete one you cannot.
Board motions
Every motion on record, picked up from your meeting minutes, and whether the follow-through actually happened. A motion with nobody answering for it cannot be closed until somebody does.
Letting it act on its own
Chief of Staff does nothing by itself until somebody grants it a mandate, and the mandate is where the limits live: who it acts for, which assistant acts, what it may read, a money ceiling, a daily action limit, and an end date.
Two things we were deliberate about:
- An amount ceiling limits the assistant. It does not replace your approval rules. Money still goes through the normal chain.
- Revoking is auditable. A revoked mandate stays on the page showing the date its authority ended, so the revocation itself can be checked later.
Agents you build yourself
Our agent library covers the work most organizations share. It does not cover the review process that is specific to yours.
Sidebar → Agents → My agents → Create Custom Agent, then Advanced, is a seven-step wizard: choose the goal, pick the module and document type, describe it in plain English, define its settings, generate and edit the prompt, test it against a sample, then set it up and turn it on.
Settings are what make it reusable
The step that matters is Settings. Each setting you define becomes a field somebody fills in when they set the agent up: a number with a range and a unit, a yes/no, a choice, a date, a state. So Finance and HR can run the same agent at different thresholds, and neither has to fork it.
Reference a setting in the prompt and it reaches the agent at run time. Reference a key you have not defined and saving is refused with a message naming the key, so the prompt and the settings cannot drift apart.
Tools carry their own permissions
Under Allowed Tools you choose what the agent may reach for beyond the document in front of it: business memory, related documents, a vendor record, a budget. Each tool carries its own permission, so granting one does not widen what the agent can see past what the people running it are allowed to see.
Guardrails that say no
Saving is refused, with the reason named, when the prompt contains a prompt-injection pattern, references a tool or variable that does not exist, runs over 10,000 characters, or uses an undefined setting.
Every agent runs at temperature zero, so the same input gives the same answer rather than a differently-worded one each time. Editing an agent publishes a new prompt version rather than overwriting the old one, and you can roll back when a change makes the answers worse.
Start from a document instead
If the procedure is already written down, Build with AI reads it. Drop in a PDF or Word document, or type a description, and Muin drafts the workflows, agents and policies it describes.
You review each one and approve it individually. There is deliberately no “Approve all”: each approval creates a real object in your organization. And everything it creates arrives inactive: an approved workflow still has to pass the review gate, an approved agent still has to be set up. Nothing imported starts running on its own.
Why an agent and not just a report
A report shows you a number. An agent watches it: runs on a schedule, checks your records, raises a task for the right role with a due date, escalates if nobody acts, and can start a workflow.
An agent that does none of that adds nothing over a report. That is the bar we hold them to.