The Raisy Blog
Multi-Agent Supervisor Architecture, Compared
J.P. Morgan runs one supervisor over three sub-agents. Raisy runs a Chief of Staff over process rooms that own state. Where the two designs diverge.
Two multi-agent systems can both be right and still not be substitutable, because the difference that matters is not how many agents each one has. It is where the human stands.
Two systems, opposite ends of one pipeline
In May 2025, two engineers from J.P. Morgan Chase Private Bank gave a talk at LangChain Interrupt about Ask David, an investment research assistant their team built for a private bank. The name is an acronym: Data, Analytics, Visualization, Insights, Decision-making. One supervisor agent sits above three specialists. One turns a question into SQL or GraphQL, one retrieves document chunks from a vector index, one calls proprietary models for fund performance, attribution and exposure. A person reviews the answer.
We built the other arrangement. In Raisy the unit is not a question, it is a process, and a process room is a real room with its own state, its own queue and its own measures. The room's caretaker is a process harness: it holds the state, runs the room's staff, enforces the budget and the tool allowlist, and reports upward. Above the rooms sits the Chief of Staff Room, whose actor coordinates and never performs. The user sits in the boardroom and may walk into any room at any time.
The identifying feature of our arrangement is one sentence: the human is not the last mile. The human is the gate from the first mile, and the machine's job is to make the gate cheap and then remove itself from it by evidence.
Teal marks ours, grey marks theirs, and blue marks the person on both sides — which is the honest way to draw it, because both designs put a human in the loop. The difference is where. On theirs the person reads a finished answer and then acts, off the diagram. On ours the person is the gate, and the act does not happen until he says so.
The last mile, and the gate
The bank's third lesson names its mechanism plainly: Human SME in the loop is the last mile of an accuracy ladder, after prompt engineering, chunking strategy, search methodology and workflow refinements have taken the system from under 50 percent to 80 and then 90 percent. The slide says of that last mile: “Hardest. May be not achievable.”
That is an honest sentence. The human is what remains when engineering runs out of road. It is also not the only place that system uses a person: for complex analytics queries it generates code and supervises the execution, which puts a human in front of the consequence rather than after it. And in May 2026 the same team published a second version that reworked the supervisor into a Deep Agent driven by the bank's proprietary private-banking Skills, described as the business's encoded knowledge of how its research work is actually done. That last move is worth remembering, because it is the direction we went in first: a kernel is exactly that, encoded once and versioned.
Four mechanisms carry the difference
- A room holds state, never memory. One company has one boardroom and one process room per process. A room owns its status, its queue items and its measures. It does not own knowledge: knowledge lives in the single entity memory, visible from the boardroom always. A stale roll-up is treated as a loud failure, not as an absence of news.
- Recall is exact, not similar. Their unstructured agent is a retrieval agent over a vector database and a hierarchical index, which is the correct design when the corpus belongs to someone else. Ours is a lookup: a fact is recalled by entity and by dimension, newest version first. Vector search has no runtime role in the memory path. The reason is not that similarity search is bad, it is that it puts variance into what the model sees, and a governed record cannot have a retrieval lottery in it.
- The asking engine, not a better prompt. Their second lesson is evaluation-driven development, and it is a good answer to a real problem. We made the question itself a governed component: never write a field that was not read from a named source, verify in a declared order before asking, never ask twice, and treat an uncited field as a schema violation rather than a guess. The question count is measured as the user's cost.
- A budget for proactivity. A system that answers is a tool the customer drives. Raisy initiates, on a clock, on an event or on a measure that moved, inside a budget whose currency is one decision waiting on the owner. Proactive work may never consume the capacity a human review needs.
The house behind the gate
Our arrangement, in the same picture language: the reporting line runs from a process harness up to the Chief of Staff and then to the boardroom, and the user's path into any room is never blocked.
The house has three spaces and each has one short objective; the objective is the test to hold a space to, so a space that stops serving its objective is the thing that is wrong. A fourth thing stands outside it, and it is not a room of this house at all.
- The boardroom — the owner's seat. Objective: be the one place where his decision is due. He is the only approver, and this is where prepared items wait, each with its evidence attached. He may walk into any room at any time, unrestricted, which is a right rather than a feature: the reporting line is a convenience, never a wall.
- The Chief of Staff Room — the reporting line, never a gate. Objective: supervise the rooms, figure it out before escalating, and curate what deserves the owner's minutes. It coordinates and never performs, so it holds no work of its own. Items do not flow through it; state roll-ups do. That is what keeps it from becoming the bottleneck it looks like it should be.
- A process room — one process, one room. Objective: run that process to completion. It holds the process's state, its queue and its measures. A room is minted when a process exists, never for an org chart. Its caretaker is a process harness that runs the room's bots, enforces a budget and a tool allowlist, and reports upward. The room holds state; it does not own knowledge.
- The gatehouse — a separate house, not a room of this one. Objective: hold the only crossing of the entity wall. It is deliberatelynot in the ladder above, and not a process room of its own: it is a peer of the whole house, with its own harness, its own failure domain and its own measures. It does not wait for a process to exist — it stands with the entity, as a wall, from the first day, and it carries two gates rather than one: what comes in, and what goes out. That separation is the point. If the thing that decides what may leave the company were a room inside the company, the same agent that wants to act would also be the one holding the door.
The Chief of Staff is the reporting line, never a wall. A room that stops reporting is an incident, not an absence of news.
What the user actually sees
On their side, an advisor asks a question and gets a curated answer, a formatted document or a chart, and a subject-matter expert has reviewed it. On ours, the user opens the boardroom and finds prepared items waiting on a decision, not an empty box waiting for a question. Each item carries its citations and can be approved, edited or rejected. Four classes reach the user and nothing else does: an external effect that needs authorising, a judgement the machine could not clear, something irreversible it cannot classify, and a loud failure with the proposed next step attached. A person who has not looked at the system for five days returns to a digest, not a backlog.
Which one you should build
The honest verdict is that these are different products, so the useful question is which one your situation is. The table below is here so a reader who reads nothing else still gets an answer.
| If this is your situation | Build | Why | Confidence |
|---|---|---|---|
| Answers, reports or charts from a document corpus you do not own | A supervisor over retrieval specialists | The corpus is the product, and a similarity index is a defensible fit for documents no one governs | High |
| A function that acts, with an external effect, for an owner | Process rooms behind a gate at the effect | The artefact is a state change, so the gate belongs where the consequence crosses the boundary | High |
| You cannot name your specialist agents before reading the company | Grow a room per process | A fixed agent taxonomy forces a design decision about a company you have not looked at yet | High |
| The same question must produce the same context twice | Exact, keyed recall | A similarity search varies what the model sees, which makes a wrong answer hard to attribute | Medium |
| Autonomy must increase over time without a policy rewrite | Earn it per unit, with the gate as an owner setting | An earned threshold is measurable and reversible; a binary switch is neither | High |
| You are regulated and keep a named expert in the loop by policy | Keep the expert, and call it a policy choice | Their own slide says the last mile may not be achievable, which may be governance rather than capability | Medium |
Where their design is the better answer
A comparison that only points one way is not a comparison. Their retrieval design is right for a corpus no single company governs, and it is what most teams should build first. Their independent evaluation of sub-agents, testing each specialist separately from the whole flow, is a practice worth copying and it is stated more clearly in their slides than in our own rules. Their metrics per agent type, including faithfulness and response relevance for a retrieval agent, are the right shape of measure for a research assistant. Our own position is narrower than a first reading suggests. We do measure the agent, but what we measure is its trustworthiness: one confidence score per bot, approvals divided by runs, and the rule that defines it separates the two instruments in as many words. We do not yet measure the agent's quality, and the reason is not a missing idea. A bot KPI is already defined as what a bot produces, output accuracy, throughput and latency. It was decided in August and its enforcement half was never built, which is a different failure from not knowing what to build. The domain itself, fund attribution and exposure, is substantive finance rather than retrieval plumbing.
They also put reflection in the flow as a first-class node, which is easy to point at. Ours is distributed across the Chief of Staff's duty to complete the evidence chain before escalating, the asking engine's declared order of checks, and the rule that a stale roll-up is loud. Distributed is more robust and harder to draw, and that is a fair criticism of how our design reads.
The gate is where the field is still improvising
One talk is not a field. The same conference and the same published collection of lessons include four other production systems, and together they say something more useful than any single comparison. BlackRock presented LangGraph agents with a plugin registry that lets more than fifty engineering teams contribute tools and agents, evaluated in daily CI; Bridgewater split its analyst tool into a chat agent and a deterministic coding agent, arguing that correctness is easier to guarantee when coding is treated as a compiler problem, and instances one agent per user scoped to that user's entitlements; Chime built its member-facing agent with legal partners co-writing the evaluations. Morgan Stanley reports that 98 percent of its Financial Advisor teams have adopted its assistant, which is the strongest adoption number in the set and a reminder that adoption and governance are different problems.
The pattern is that everyone has converged on the same agent shapes, and almost nobody has converged on where the human stands. J.P. Morgan's own written guidance puts it as a principle rather than a feature: governance and compliance move into runtime, with spend limits, approval checkpoints and redaction treated as runtime concerns rather than review-time paperwork. We agree with all of it, and we took it one step further. If governance belongs in runtime, then the gate is not a screen bolted onto a finished pipeline. It is the reason the pipeline has the shape it has.
Why this matters beyond convenience
Every serious agent deployment eventually meets the same question: what does the system do when nobody is watching it. The two architectures answer it in opposite ways. One keeps raising accuracy until a person can stop checking, and where that is impossible, keeps the person. The other treats the person as a permanent gate and treats the machine's progress as the gate getting quieter: an approval rate that climbs, an override rate that falls, a rung of autonomy earned per unit of work. Both can be honest. Only one of them is a claim about the owner's time.
That is the promise worth holding the system to: AI prepares, humans decide, trust is earned. A gate that never opens is a bottleneck with good manners. A gate with no evidence behind it is a liability. What makes the arrangement work is that every crossing is recorded, every authorisation is attributable, and the width of the gate is set by the owner rather than by the builder.
Open your company at raisy.team/entities to see what a run company looks like from the owner's chair.
FAQ
Is a supervisor agent the same thing as an orchestrator?
A supervisor agent is an orchestrator that also decides. It receives the question, plans the route, chooses which specialist to engage and reflects on the result. An orchestrator that only routes adds a hop. The distinction matters because the supervisor is where the judgement lives, and in a governed system judgement is the thing that has to be attributable.
Why does Raisy not use vector search for memory?
Because memory is a governed record, and a similarity search would vary what the model sees between two identical questions. Facts are recalled by entity and by dimension, newest version first, which is a lookup rather than a search. Vector search is still permitted at design time, where a choice is made once.
What is a process room, in one paragraph?
A process room is a real room inside one company, one per process, holding that process’s state, its queue and its measures. Its caretaker is a process harness which holds state, runs the process’s bots, enforces a budget and a tool allowlist, and reports upward. The room does not own knowledge; knowledge is shared across the company.
How does the gate open without a human reading everything?
By evidence. Each unit of work carries a record of approvals and overrides, and the width of the gate is a setting the owner holds: review every run, review before anything external, or let anything above a stated confidence pass. The approval mode resolves from the narrowest level outward, so a step can be stricter than the process it belongs to.
Do the two systems compete?
They can, and they can also be layered. A research assistant is a good consumer of answers; an operating system is a good place to put the gate. The mistake is to buy the first while needing the second, then discover the missing piece is not another agent but a boundary where a consequence can be authorised.
What does "stale roll-up is a loud failure" mean in practice?
It means a room that stops reporting is treated as an incident, not as an absence of news. If a room’s state is stale and no one can see it, the room above it is blind, which is the exact failure that shared visibility exists to prevent.
Does Ask David publish accuracy or adoption numbers?
It publishes an accuracy curve, under 50 percent to roughly 80 and then 90 percent, with the last mile left to a human reviewer, and a scope statement counting thousands of investment products. It does not publish user counts or production usage metrics. The large numbers sometimes attributed to it, 200,000 users in eight months and 250,000 employees with access, belong to that firm’s separate internal assistant, which is a different system.
Is a Skill in their second version the same thing as a kernel in Raisy?
They solve the same problem from opposite ends. A kernel is the unit we version: a piece of prompt or knowledge with an identity, an owner and a history, assembled into bots and processes. Their Skills are the same instinct, the bank’s encoded knowledge of how its research work is done, attached to a supervisor that became a Deep Agent in 2026. The difference is what surrounds the unit: ours sits inside a house with rooms that hold state and a gate at the external effect.
Sources
J.P. Morgan Chase Private Bank, "How JP Morgan Built An AI Agent for Investment Research with LangGraph", LangChain Interrupt 2025, published 29 May 2025 (presenters David Odomirok and Zheng Xue): the architecture, the three lessons and the accuracy ladder. · The same team's architecture slides, shared with us and read 7 October 2026: The Approach, the two sub-agent descriptions, Graph: Workflow, and Lessons 1 to 3. · ZenML LLMOps Database, "Multi-agent investment research assistant with RAG and human in the loop", read first-hand 7 October 2026: the DAVID acronym, MongoDB as a storage layer, the text-to-code execution gate, the personalization, reflection and summarization nodes, and the development order. · Farhan Baluch, "Inside JPMorgan's Investment Research Agent Architecture", 30 May 2025, read first-hand the same day: reproduces the diagram and the three lessons. · JPMorganChase, "LLM Suite wins American Banker's Innovation of the Year Grand Prize", read first-hand: 200,000 onboarded users in eight months, a different system from the research assistant. · CNBC, 30 September 2025, read first-hand: 250,000 employees with access and roughly half using it daily, also a different system. · BlackRock, "How BlackRock Built Production AI Agents for Asset Management with LangGraph", LangChain Interrupt 2025, published 12 June 2025: URL verified, and the operational detail in the peer-set section comes from secondary write-ups of the same event, which this post discloses in "What this post does not answer". · Raisy's own architecture of record, cited by rule and decision ID rather than restated: the manifest rules on the human gate, runtime recall, the isolation boundary, the governed substrate and proactivity, and the project files carrying the Chief of Staff doctrine, the asking engine and the proactivity budget. Nothing in this post is a measurement of ours and no endpoint call backs a number in it.
Keep reading

An OpenAI Dots Alternative for Companies
Dots gives one person their own computer. A company needs one owner for a function, a gate before anything leaves, and a number somebody watches.

Raisy Closed Beta Is Open: Fundraise and Sell at Once
Raising takes months you should be spending on sales. Raisy does the toil on both, and you keep the decisions. The gates close at the end of September.