Ten acquisition agents, inside authority

InsightRollup's ten agents run in order on one pipeline when someone presses Analyze. Each runs as itself, at its own autonomy level, reads the engines' facts for its part of the playbook, reasons over them with the model when one is configured, and prepares at most one decision for a person. None of them is the source of a number.

Three steps, ten agents

Agent	Step	What it reads and finds	What it may prepare	Never invents
Book intake	Book review	Data room against the checklist; data quality; company profile, leadership and news from InsightPulse	Ask the seller for missing documents and corrected data	A document or row the seller didn't send
Retention analyst	Book review	Client and revenue retention by producer and cohort; organic growth	—	A retention rate
Concentration analyst	Book review	Carrier, producer and client concentration; key people; producer agreements	Propose retention agreements for key producers	A share
Revenue quality analyst	Book review	Recurring vs contingent; renewals in 90 days; book against the seller's revenue	—	Revenue or a price
Cross-sell analyst	Book review	Single-line commercial clients; the acquirer's lines they don't buy	—	— (the estimate is the engine's)
Client match agent	Integration fit	The target's clients against the acquirer's: exact, website, Pulse, name	Confirm the proposed client matches	A match
Carrier mapping agent	Integration fit	The target's carriers against the acquirer's appointments	Request carrier appointments	—
Integration planner	Integration fit	The findings and engine figures	Adopt the 100-day plan as tasks	—
Valuation memo writer	Memo & controls	The approved findings	Finalize the valuation-support memo	A figure
Supervisor agent	Memo & controls	The acquisition controls	The practice leader's sign-off	—

The agent catalogue: ten agents across book review, integration fit, and memo and controls, each with the action it may propose and what it never invents.

One run

sequenceDiagram participant U as Deal lead participant P as Pipeline participant E as Engines participant A as Agent (each of ten) participant M as Model participant G as Grounding check U->>P: Analyze loop intake, retention, concentration, revenue, cross-sell, client match, carriers, planner, memo, supervisor P->>E: Facts for this agent's playbook items (with citations) E-->>A: Numbered facts F1..Fn A->>M: Reason over the facts (fixed answer shape) M-->>G: Summary, findings with fact ids, open questions G-->>A: Kept, or thrown away if a number is not in the facts A-->>P: Findings, at most one prepared decision per action end P-->>U: Narrative, findings, prepared decisions

Each step is stored: the agent, its level, whether the model ran, the model, tokens, cost and how long it took. A prepared decision supersedes any still-open one for the same agent and action, so there is one live ask per action.

An agent's runs and one run's trace — The concentration analyst's runs and one trace, captured with the model switched off: what the book said, the step that reasoned (here by rules), and what was prepared for a person.

The trace above was captured with the model switched off, which is a useful thing to see: the agent's own engine findings stand, the trace says so in words, and what is prepared is unchanged. With a model configured the middle step shows the model, tokens, cost and the reasoning, all grounded in the same facts.