Holding the Whole and Doing the Detail: Three Layers of AI in a Small Team

On 9 October, in the middle of handing out work, I asked the system I work with a very ordinary question: if the coordinating side runs out of its weekly quota before the reset date, who carries on? Weekly quotas are a real constraint of AI tools, not a theoretical one. A company that leans on a single source stops dead at exactly the moment work is running.

That question forced me to state plainly how HOPE organises its work. I am writing it down because it answers something investors and partners often ask: how does a small team operate like a large one and still stay in control?

Three layers instead of one "AI that knows everything"

I do not believe in a single agent that does everything. HOPE's structure has three layers, much like an ordinary company with executives, department heads and the people who do the work.

• The top layer is one coordinating agent. It sees the whole company, takes in goals, splits work across the departments and gathers the results.
• The middle layer is a management agent for each department. The company is divided into nine departments, including Brand and Communications, Finance and Fundraising, Operations and Product, and People and Culture.
• The bottom layer is small agents, each with one detailed job: parsing order messages into the revenue ledger, drafting replies to customers, computing monthly payroll, drafting posts and passing them through a banned-claims filter.

With this structure, a team of 12 people (9 full-time) has automated more than 60% of its work. I do not lead with that number, because a count only means something when it sits next to controls. What I want to stress is the shape.

My original pain was that departments did not know how to work with one another. The middle-layer agents exist to connect departments proactively, instead of waiting for a person to stand in the middle and relay messages. On maturity, my own assessment is that the structure is past 60 percent: it runs fairly smoothly and is not yet at 100. That is a self-assessment, not an audited figure.

Holding the whole picture and doing the detail are different jobs

I told the system that one side should keep hold of the overall work while the other handles the detail. I split them because the two jobs demand very different ways of working.

Holding the whole picture needs wide context: knowing what takes priority, remembering which commitments are open, noticing where two departments are stepping on each other. Doing the detail needs focus: going deep into one line of work, running for a long time, correcting until it is right. Put both into one agent and it has to keep the big picture while sinking into detail, and its context fills up until it forgets the important thing.

Split them and each side keeps a clear head. The coordinating agent is not pulled around by detail. The executing agent works within a narrow scope, so when it is wrong it is wrong inside a small compartment. There is also a benefit any manager will recognise: the one who decides is not the one who does. The doer does not mark its own homework, so risk is controlled by structure, not only by instructions.

The bridge is a work queue and a status report

The two sides do not chat freely. Work handed down is a written brief placed in a queue. Results come back as a report with a status that takes only three values: done, partly done, blocked.

Every brief carries its own rules. Do not delete anyone's data, do not message customers, do not expose secrets. Where something is uncertain, stop at that item and write it in the report instead of guessing. If it cannot be done, say so and say why; never pretend it is finished. This connects directly to the principle I wrote about earlier, AI reads wide, writes narrow: the narrower the write permission, the safer a multi-layer structure becomes.

Three statuses sound crude, but that is exactly why a reviewer reads them quickly, and why the machine has no room for a vague "done".

Cross-cover when one side runs out of quota

Back to the opening question. The executing side is divided into departments that mirror the coordinating side. So when one side runs dry, the other can step into its role: it already knows the department structure, and the queue shows which work is half finished and in what state. I want the system to take over proactively when the other side has run out of quota before its reset, instead of standing by and waiting.

I will be direct about the limit. Cross-cover is a design requirement I set, not an absolute guarantee. Stepping in does not mean replacing everything. The substitute keeps running work from breaking, while major decisions still wait for a person to approve. A system with a single point that can fail is not an operating system; it is a weakness on a timetable. I would rather accept some overlap between the two sides than let the company depend on one of them.

Where the people are

I told both the system and the team that I would stand in a supervising role. The process I want is a loop: agents survey, propose a strategy, I approve, work is carried out, results are reported, and when a problem appears they propose again. I should not be the person who touches everything by hand.

Supervising does not mean absent. Three places stay permanently with people.

• Everything that goes out to customers or the public. The machine drafts; a person approves before it is sent.
• Money and anything irreversible. AI has no authority to move money, delete data or change infrastructure. Advertising spend is off by default, with a cap of VND 1 million per day (about US$38).
• Direction. Choosing goals, and choosing one option from the two or three the system proposes.

On top of that, four automated evaluation suites run every week, so the quality of the agents does not drift without anyone noticing.

What partners should understand about HOPE

First, our costs grow with software, not with headcount. That is why a small team can carry the volume of work of a large one: repetitive and detailed work sits in the bottom layer, and people stay with what is hardest to replace, which is judgement, relationships and accountability.

Second, the hardest part is not writing the agents. It is dividing roles so that departments fit together: who decides, who does, who checks, and in what format they report. We are still tuning this every week.

Third, I do not hide risk. Our own records list dependence on the founder as a real risk. The three-layer structure, together with written processes and weekly advisors, is a way to reduce that risk, not to erase it. My long-term goal is a company that can be run by very few people, with even the application work done alongside AI. That is a direction, not the present state.

If you are considering working with HOPE, the question I hope you ask is the one I asked myself: when one link fails or runs out of strength, does this company stop, and who knows?