Ten placement agents, inside the broker's authority

The agent catalogue: the renewal and submission agents and the markets and quotes agents, each with its action and what it never invents — **The catalogue.** The agent catalogue: the renewal & submission and markets & quotes agents, each with its action, the level it starts at and what it never invents.

InsightPlacement has ten agents. Each has one job in the placement lifecycle, reads a defined set of facts, prepares one or two kinds of decision, and stops there. None of them sends, selects or binds anything by itself: what goes to a client, a carrier or a binder is a person’s call.

The pipeline

Analyze runs the agents in order on the shared agent pipeline, each as itself at its own autonomy level. Each step gathers facts from the engines, reasons over them with the model (Claude Haiku 5.5 by default for the agents), passes the grounding check, and either prepares a decision for a person or records why it had nothing to do (“Not yet: markets are chosen once the submission is approved”).

flowchart TB subgraph P1["Renewal and submission"] direction LR A1["Renewal
readiness"] --> A2["Data-gap
chaser"] --> A3["Submission
builder"] end subgraph P2["Markets and quotes"] direction LR B1["Market
selector"] --> B2["Outreach
drafter"] --> B3["Quote
extractor"] --> B4["Quote
comparison"] end subgraph P3["Proposal, compliance and bind"] direction LR C1["Proposal
writer"] --> C2["Placement
compliance"] --> C3["Supervisor"] end P1 --> P2 --> P3

Each agent, what it reads and what it prepares

Agent	Reads	Prepares	Never invents
Renewal readiness	The 120/90/60/30 calendar, exposures, loss runs, the data room	Order loss runs from the incumbent carriers	A date or a loss figure
Data-gap chaser	The statement of values, exposure data	Ask the client for the missing data	A value the client didn’t give
Submission builder	Client record, SOV, loss runs, expiring policy, InsightPulse profile	Approve the submission as market-ready	A figure — every one is the engines’
Market selector	The appetite table, the broker’s results, appointments	Adopt the marketing plan	An appetite or a rating
Outreach drafter	The approved submission, the plan	Send the submission to the planned markets (a dry run)	—
Quote extractor	Quote letters in the data room	Confirm the extracted quotes	A premium or a term the letter doesn’t state
Quote comparison	Quotes, the expiring program, carrier ratings	Select the recommended quote for each line	A price or a ranking
Proposal writer	The selected quotes, the comparison	Finalise the client proposal	A figure
Placement compliance	The declinations log, carrier ratings	File the surplus lines diligent-effort affidavit	—
Supervisor	The placement controls	A practice leader’s sign-off when a control fires; the bind order when all is clear	—

Gather, reason, prepare — and what a person sees

sequenceDiagram participant E as Engines participant A as Agent participant M as Model participant G as Grounding check participant P as Person E->>A: numbered facts, each with its source A->>M: facts and the question for this step M-->>A: findings in a fixed shape, citing fact numbers A->>G: every number in the answer must be in the facts G-->>A: pass, or throw the answer away A->>P: prepared decision: reasons, facts, what it could not settle P-->>A: Accept, Reject or Ask, signed on the record

Three things are always on a prepared decision card: the reasons (bullets a person can read in ten seconds), what the agent couldn’t settle (its open questions, so a person knows where to look), and the facts behind it (numbered, from the engines). The card also says whether the agent reasoned with the model or stands on the engines alone.

The compliance agent's prepared affidavit with the five facts behind it, on illustrative data — **The facts behind it.** The facts behind a prepared decision (illustrative): the surplus carrier and the three admitted declinations, each with its reason — numbered, from the engines.

Every run is traced

One market selector run's trace on illustrative data: status, level, model, tokens, cost and time, and the three steps — **A run, traced.** One market selector run, traced (illustrative data): what the engines said, the model step with its model, tokens, cost and seconds, and what was prepared for a person.

Each agent’s page lists its recent runs with the model, tokens, cost and time, and a trace for each run. On the illustrative book, one market selector run used about 1,700 input and 2,200 output tokens and cost a fraction of a cent; most runs needed no model call at all because there was nothing new to reason about.

For what stops an agent from doing more than prepare, read the autonomy dial; for why its numbers can be trusted, grounded figures.