MEET CLAUDETE
The closest thing I have to an office manager.
Inbox triage · Calendar coordination · Daily report digest · Controlled outbound communication
Six incoming reports
↓ Sorting layer: urgent · deadline · waiting · monitor
↓ One prioritized list, capped at eight items
Six reports become one decision-ready morning view.
Six is the number of incoming report sources; eight is the maximum length of the resulting list.
Claudete is not a chatbot attached to a spreadsheet. She has her own inbox, not simply access to mine.
Each morning, she reviews what has arrived, identifies what is urgent, and brings it forward the same day. If something requires a reply and I go quiet, she follows up once—enough to help, never enough to become noise.
She manages my calendar so scheduling does not become one more task competing for attention.
By 8:00 a.m., she has also turned the other agents' reports into one prioritized list, capped at eight items and ranked by actual deadlines. I do not read six reports before coffee. I read one.
Claudete is currently the only agent permitted to send anything on my behalf. Every other agent may recommend, draft, or flag. She is the only digital voice allowed to leave the building—and even that role operates within defined boundaries.
Those boundaries are specific. Claudete's outbound actions—triage, scheduling, one follow-up nudge—run inside rules I have already approved, so they don't wait on a live decision each time. That is still Monica-in-the-loop: every action is reviewable, revocable, and logged, not autonomous in the open-ended sense. She has no access to money. She cannot see or use a bank account, a credit card, or any payment method, and won't until that safeguard has been tested as carefully as every other one in this system.
No agent, including Claudete, holds financial access.
ONE BOTTLENECK AT A TIME
Each agent began with work that was already falling through the cracks.
The first problem was a pile of receipts I never had time to log. That became the Finance Agent.
Then development moved from front-end design into backend work, and I needed a clearer view of progress and risk. The Prototype Manager followed.
The need for capital introduced another recurring workload: finding opportunities, preparing submissions, and tracking what I still needed to complete. That became the Grant Agent.
As the system expanded, I needed a way to review the reviewers—to see what was useful, what was drifting, and where my attention was needed. That became the CEO Review Agent.
The goal was never to create as many agents as possible. Each role had to earn its place by reducing a specific form of friction.
Finance Agent
- Triggering friction
- Unlogged financial records
- What it prepares
- Organized financial information
- What remains human
- Review and financial decisions
Prototype Manager
- Triggering friction
- Fragmented progress reporting
- What it prepares
- Progress, risks, and follow-ups
- What remains human
- Priorities and direction
Grant Agent
- Triggering friction
- Scattered funding work
- What it prepares
- Research and submission preparation
- What remains human
- Opportunity selection and approval
CEO Review Agent
- Triggering friction
- Too many agent outputs
- What it prepares
- Cross-system assessment
- What remains human
- Final judgment and intervention
THE WORK WAS NOT CLEAN
The useful lessons came from the friction.
The early workflows did not arrive fully orchestrated. Some produced too much information. Some needed clearer priorities. Some needed stronger limits around what they could recommend or send. Connecting multiple roles created new questions about duplication, escalation, accountability, and trust. The system improved because we treated those moments as design evidence—not as reasons to hide the experiment.
| Early behavior | Design response | Improved behavior |
|---|---|---|
| Too much output | Cap and prioritize | Decision-ready summaries |
| Unclear handoff | Assign one owner | Visible accountability |
| Trust assumed too early | Add review stages | Responsibility expands gradually |
What agents can do—and where they stop.
| Agents may | A human must |
|---|---|
| Organize information | Set direction |
| Compare inputs | Interpret consequences |
| Draft responses | Approve outbound communication |
| Flag risk | Accept or reject risk |
| Recommend next steps | Make consequential decisions |
One narrow, named exception: Claudete's routine, pre-approved actions (inbox triage, scheduling, a single follow-up nudge) run inside rules Monica has already set, so they don't wait on a live decision each time—see Meet Claudete. Everything else, including all financial activity, still crosses the line above.
The goal is not autonomy for its own sake. It is clearer information, better-timed decisions, and more room for the work that requires human judgment.
Useful AI needs boundaries, ownership, and evidence.
- 01
Start with the work, not the technology.
An agent should solve a real operational problem. Novelty is not a business case.
- 02
Recommendation is not authority.
Agents may propose a next step. Consequential decisions remain human decisions.
- 03
Give every decision an owner.
Someone must remain accountable for what moves forward, what stops, and what leaves the company.
- 04
Measure before expanding trust.
We review usefulness, accuracy, and behavior before allowing a workflow to take on more responsibility.
- 05
Design the checkpoint before the shortcut.
Speed is valuable only when the review process protects the people and work affected by it.
- 06
Share the lesson. Protect the implementation.
We can explain what the system teaches us without publishing the instructions and access that make it operate.
WHAT OUR LAB IS PREPARING US TO DO
Small businesses do not need more AI noise. They need a place to begin.
Many founders can see that AI may help their business, but the path from curiosity to a responsible operational workflow is rarely obvious. Which problem is worth solving first? What should remain human? How should outputs be reviewed? When has a workflow earned more trust? Our Lab is how Forma Labs is developing informed answers: by implementing the work internally, measuring where it helps, documenting where it fails, and refining the boundaries around it. We are not offering agentic-operations consulting yet. We are building the experience required to offer it responsibly in the future.
Where would you begin?
Where small business can leverage agentic AI.
Manage emails
Sort messages, draft replies, and follow up with customers.
Track finances
Categorize expenses, organize receipts, and flag overdue invoices.
Create marketing content
Draft social posts, newsletters, and website updates.
Support customers
Answer common questions and route requests to the right person.
Coordinate schedules
Arrange meetings, manage calendars, and send reminders.
Follow up on sales
Organize leads, prepare proposals, and track conversations.
Manage daily operations
Assign tasks, monitor progress, and identify delays.
Review business performance
Summarize reports, spot trends, and prepare information for decisions.
ASK MONICA
Where is work getting stuck in your business?
Our Lab is still evolving, and we are not offering agentic-operations consulting yet. But I want to hear where founders and small teams are losing time, repeating work, or carrying decisions that never seem to reach the top of the list.
Today, if you ask me something here, it comes straight to my inbox. Someday, I suspect Claudete answers first.