An asset buyer's agent-first strategy, drawn as a system: what to build first
A client that acquires assets at volume started from AI-assisted operations: analysts doing the work with an AI assistant. NoArt Systems helped it move to an agent-first way of working — agents research, assemble and propose; people confirm — and drew that strategy as a system of two engines and ten feedback loops, so the roadmap could be ranked by structure.
The same structure appears wherever agents feed a human approval step: unless approval capacity rises with agent output, more agents mean a longer queue, not more decisions.
Generalized from one client engagement. Names and figures are removed.
Situation
Sequencing the roadmap needed a model of how the strategy would turn.
The client is a specialist investor that discovers, evaluates and acquires assets recorded in a large external record, with an analyst team approving every acquisition. It started from AI-assisted operations: analysts did the research and assembly themselves, with an AI assistant at their side.
Working with NoArt Systems, it consolidated that record internally, gave software agents a way to read it, and moved to an agent-first way of working: agents research, assemble and propose; people confirm.
The open question was sequencing: which feedback loops the strategy depended on, which of those already existed, and which limits would stop it.
Three terms carry the argument. Write-back is the act of putting what an agent or an analyst learned about an asset back into the shared record, so the next reader starts from it. Record depth is how much is known about each entry in that record. The desk is the analyst team whose approval every acquisition needs.
Approach
Draw the strategy as a stock-and-flow system, then check each reinforcing loop against what already exists.
NoArt Systems modeled the business as stocks (assets at each stage, record depth, tool coverage, capital), the flows between them, and the feedback loops those flows close. Each reinforcing loop was then checked link by link against what already exists, to find where a loop would stay open.
NoArt Systems built the map into the client's own tooling as an interactive figure, beside the live measurements of the record it describes, so the model and the data it makes claims about sit on one screen. Every number on that screen carries its measurement date; the diagram carries none.
The map, generalized
Seven stocks, thirteen flows, seven reinforcing loops, three balancing loops, four policy levers. Step through the chapters; on a desktop, hover any box to trace its links.
Scroll the figure sideways to see all of it.
Chapter 1 of 8: The agent-first flywheel
Figure W1. The agent-first flywheel on a unified record. Teal boxes are stocks, valves are flows, gold labels are policy levers. A blue plus marks a link in the same direction; a red minus, one in the opposite direction. Generalized from client work.
All eight chapters and ten loops, as text
- The whole structure
The agent-first flywheel
Two engines — context (R1 · R4 · R5) and prediction (R2 · R3 · R6) — plus capital recycling (R7), held back by three balancing loops (B) and by how fast the desk can decide, on one unified record.
The trajectory is set by the feedback structure, not by the targets: change the loops and the wheel turns. Step through the chapters to light each one; on a desktop, hover any box to trace its links.
- The grammar
Stocks, flows, loops, levers, delays
Only flows can be managed; stocks follow. The leverage is in the loops.
Stocks accumulate — they are the system's memory, and the reason nothing changes overnight. Flows fill and drain them. Reinforcing loops (R) compound; balancing loops (B) resist. A flywheel is a set of R loops whose B loops you have deliberately pushed outward.
Levers are the exogenous inputs a decision actually sets: agent capacity, analyst headcount, sources on the sweep, capital injections. Delays (‖) make a policy look wrong, or right, long after it was set.
- The physical pipeline
Seven stocks, thirteen flows
Everything the business does is a flow between these stocks — and the desk sets the pace of the one that matters most.
From discovery to portfolio: untapped assets enter the queue, the desk decides, negotiation closes, the contract is signed and the outcome pays — returning cash to capital. Discards and drop-outs leave by their own outflow.
The desk is the governor: approval runs at analyst decision capacity — headcount, a lever, times per-analyst productivity, which agents can raise. If agents fill the queue faster than the desk empties it, acquisitions plateau at desk capacity and the excess leaves as discards and give-ups; counterparties also drop out in negotiation.
- Engine 1
Context compounds (R1 · R4 · R5)
Agents are only as good as the record behind each call; the loops close only if diligence and outcomes write back to it.
R1 — one call returns the whole record: fields, dates, related records. R4 — the tools agents can call, drawn as a stock: demand exposes the next gap. R5 — if agents pre-assemble the decision pack, each decision becomes a confirmation.
The write-back path carries this engine: until diligence findings and outcomes are written back into the record, the context engine is an open arc.
- Engine 2
Prediction pays for depth (R2 · R3 · R6)
A sizing model and an outcome model turn record depth into target precision; outcomes and cash flow back into the record.
R2 — better targets raise yield, which brings acquisitions, which return cash, which buys depth; the cash leg carries the longest delay. R3 — every outcome is a label, once written back. R6 — models surface theses across the whole record, including pools a target list would miss.
Depth is worth buying only where the model can use it: sample a third-party source before buying it in bulk.
- Limits, and the escape
Finite pools, visible market (B1 · B2)
Thesis pools are finite and buying is visible; discovery across the whole record is what moves the limit outward.
B1 — as proposals drain the pool, yield falls; the portfolio grows S-shaped on a fixed pool. B2 — acquisitions become visible to the market and competitors arrive after a delay. The way out is refilling the pool: a wider sweep, and discovery across the whole record (R6, in Engine 2), find pools a target list would miss.
- Where the money is
The fuel tank (B3 · R7)
A business buying assets that pay out slowly has a fuel tank — and a tap: capital injections are the fourth lever.
B3 — every contract and depth purchase ties up capital that returns only when the asset pays out; where capital is scarce, this limit governs volume. R7 — assets that pay out refill the tank, with the longest delay in the system.
The agent engines decide which assets to buy and how fast the desk says yes; capital can cap how many.
- The decision
What to build first
Three leverage points, in the order they make the flywheel turn — funding sets how big it can get.
1 — Write-back discipline: every diligence finding, every document reviewed, every outcome lands in the record. Without it R1, R3, R5 and R6 are open arcs.
2 — Agent-assembled decision packs (R5): raise decision capacity before proposal volume, or the queue turns into a backlog and the excess leaves as discards and give-ups.
3 — Refill the pool, fund the tank: point the sizing model at the whole record and widen the sweep (R6), and set the injection schedule for capital (the fourth lever), which can bound volume whatever the agents do.
- R1 · Context compounds
- One agent call returns the whole record for an asset — the agent researches, drafts and proposes from it. The loop closes only when what the agent learns is written back.
- R2 · Prediction pays for depth
- Better targets → more acquisitions → cash from payouts → bought depth in the record where the money already showed.
- R3 · Outcome rate trains the models
- Every outcome is a training label — once it is written back into the record. It arrives late, well after the acquisition.
- R4 · Tooling grows with use
- Agent demand exposes the next missing tool; each tool added raises the context of every later call.
- R5 · Agents widen the desk
- If agents pre-assemble the decision pack, the analyst's decision becomes a confirmation, not an investigation. The loop the whole structure leans on most.
- R6 · Models surface theses
- Discovery runs on breadth — how many sources are being scanned — sized by a model over the whole record. It finds pools a target list would miss.
- R7 · Capital recycles
- Assets that pay out refill the tank — with the longest delay in the system. Real, but slow.
- B1 · Thesis exhaustion
- Each thesis pool is finite. As proposals drain it, the yield per counterparty contacted falls.
- B2 · Crowded market
- Acquisition becomes visible to the market; competitors reach the same counterparties after a delay.
- B3 · Capital tied up
- Every contract and every depth purchase ties up capital that returns only when the asset pays out. Where capital is scarce, this limit governs volume.
Terms used in the figure
- Stock
- An amount that accumulates over time — assets at a stage, capital, how much is known.
- Thesis
- A rule for which assets to go after. Each thesis defines a pool.
- Desk
- The analyst team whose approval every acquisition needs.
- Record depth
- How much is known about each entry in the shared record.
- Depth purchase
- Buying third-party data to deepen the record.
- Write-back
- Putting what was learned about an asset back into the record for the next reader.
- Sweep
- The set of external sources being scanned for new entries.
- Record breadth
- How many external sources the sweep covers; a policy lever.
- Intake
- Getting a proposed asset into the decision queue; the intake queue holds those awaiting a decision.
- Sizing model
- Estimates the size and value of a pool, or of one asset, from the record.
- Agent context
- What an agent is given to read before it acts.
- Outcome
- How an acquired asset ends: it pays out, or it is written off.
- Outcome model
- Estimates how an asset will end.
- Counterparty
- The current holder of an asset, who sells it (shortened to "party" in the diagram).
- Proposal
- An agent's write-up of one asset, for the desk to decide on.
- Decision pack
- Everything the desk needs to decide one proposal, assembled before the decision.
Deliverables
What NoArt Systems handed over.
- D1 System-dynamics map of the strategy Stocks, flows, loops and levers, with the links each reinforcing loop depends on.
- D2 Interactive figure in the client's tooling A chapter-by-chapter tour beside the live measurements, with an integrity test that model and artwork stay in step.
- D3 Measurement-date rules for the client's internal screen Every figure carries its measurement date; the diagram carries none.
- D4 Ordered leverage points Write-back first, agent-assembled decision packs second, breadth and funding third.
What the map ranks
The order the structure implies.
The map gives a roadmap a shared object: a loop, and whether its links exist. It ranks the links to build in the order they make the wheel turn. In this structure the write-back path comes first, because four of the seven reinforcing loops close through it. It also comes before adding agent capacity: more agents raise proposal volume, and until decision capacity rises with it, that volume turns into a backlog and the excess leaves as discards and give-ups.
Tell us what you are building — or what you are trying to figure out.
In your note, include the initiative, the decision point, the timeline, and whether you need advisory, architecture review, vendor oversight, or roadmap support. We respond directly.