A client asks a strategy question and gets back senior-consultant work — reviewed adversarially, trained automatically — from a system that costs a fraction of the model it competes with. Oli is the AI platform I built at More is More: eleven specialist agents forced through the right reasoning protocol.
The first version of the agents defined what to produce and not how to think. The result was predictable and instructive: an economical model with no reasoning instruction produces output that looks like consulting and is decorative — the shape of an answer without the work behind it.
An output contract fixes the format. It does not make the reasoning happen. That was the finding that redirected the whole product.
Every agent runs a shared adversarial protocol before it is allowed to answer. It is the sequence a senior consultant performs by instinct and a cheap model never performs unprompted.
Break the question into the pieces that can actually be checked.
Each part validated on its own before anything is assembled.
The step that separates analysis from opinion — and the one no model does unless you make it.
Not "87% confident". Confidence broken down by claim, because a single number hides exactly where the answer is weak.
What happens after the recommendation is taken — the question that distinguishes a consultant from a summariser.
The trade-off is real and I would state it in any pitch: five extra reasoning steps cost tokens and latency. A cheap model doing the protocol can take twice as long. The bet is that a client prefers waiting eight seconds for a correct answer over three for a decorative one.
Without this, each correction dies in its own chat window. With it, the platform gets better at the rate its users use it — which is the only kind of moat a small consultancy can actually build.
A correction inside the conversation is captured as a training candidate rather than discarded.
Candidates processed hourly instead of in real time — a single correction should not move the system.
A trainer agent looks for repetition across corrections and proposes skill changes only where a pattern exists.
Skills are files: versioned per change, rolled back by symlink. Knowledge lives in markdown; the database holds only operating state.
The hard boundary in that loop: a correction specific to one client must not contaminate a skill shared by all of them. Distinguishing "this agent was wrong" from "this client works differently" is the design problem that makes or breaks multi-tenant agents.
The platform sits behind its own AI gateway, which routes each task to a capable, automatic or cheap tier depending on what the task actually needs — and removes the dependency on any single provider's credits.
specialist agents, each with its own skill set
Tiered by task, so the expensive model is used where reasoning depth changes the answer and nowhere else.
No skill reaches production without being tested on a real case. No version is ever deleted.
Any claim about an agent's quality is backed with real output, never with impressions.
The commercial question underneath Oli is one I have answered before in other categories: where does the value actually sit, and is anybody charging for it? Here it sits in the reasoning protocol, not in the model — which means the moat is transferable across providers, and the cost base stays low enough for a consultancy to price the product for a mid-market client.
An agent that hallucinates less than a general-purpose chatbot already justifies its price. Eleven of them with adversarial review and a training loop is a defensible position. That is a positioning argument, not an engineering one, and it is the reason this sits in my portfolio rather than in somebody's repo.
Specialist agents with a reasoning protocol, reviewed before they answer.
Automatic training from the work itself.
Economical models, made rigorous.