# Make Your Agents Insurable — And The First Fact This Pivot Would Produce

**version** v0.33.80
**date** 31 August 2026
**from** Human (project lead)
**to** Strategy, the RiskMandate team, the registry site

**type** Strategy brief — memo 9, first of two beyond the original eight

*Produced from a voice memo of 31 August 2026, carried verbatim below and then read against the corpus by the site agent. Everything in the transcript is the project lead's; everything under the reading is the site agent's and is labelled as such. **The memo series recorded as eight has grown to ten**, and the folder's counter — which is a claim — has been corrected rather than quietly adjusted. This memo is also the first in the series that would produce an **external fact** rather than a position, which is noted in §1 as the thing the pivot most needs.*

---

## What This Is

The positioning, and a research programme underneath it: **the memo proposes "make your agents insurable" as the tagline for what this work is, on the grounds that it is honest about the state of the market, actionable once the data exists, and doubles as market research — because to say what makes an agent insurable you first have to establish what insurance for agents exists today, whether a Claude Code session can be insured at all, what it would cost, what it would cover, and what the requirements are; it asks the inverse question that defines cover more sharply than the cover does, which is what a holder can do that breaks it; it wants the materials mapped as well as the answers — whether an application is a web form, an API, a spreadsheet or a Word document, and how evidence is actually supplied; and it frames the whole thing as a pulse that can be taken repeatedly as the market evolves, showing where the gaps are and where this project can help.** It is the sixth document of 31 August (cross-ref: v0.33.71–79, `measure.py` and its evidence classes, GM-D40's two channels and GM-D57's connectors). New contributions: **the positioning, the breaks-the-cover inversion mapped onto the estate's own model, and the survey specified as a dated re-runnable measurement rather than a report.**

## The Memo, Verbatim

*Transcribed by otter.ai; carried whole, exactly as received.*

> So here's a interesting concept for us to tackle, which is the idea of what it's needed to make agents insurable. So almost like the tagline is "make your agents insurable, which is quite an interesting angle, because it also allows us to do a nice market research, and also publish the materials and publish the requirements, which basically allow us to say, right today, you know, August 2026 or early September 2026, what insurance exists for agents? Like, can you insure Claude? Can you insure you know my Claude code session? Can you insure this session? And how much does it cost? And what does it cover? And what's the requirement to do that? Like, is it just off the bat? You know, like do what, what will you, what can you add? Because I think what's also interesting when you talk about insurance is what can you do that breaks the insurance? What can you do that breaks the insurable part of the insurance? Right. So you know, making the agents insurable is actually a really nice Way to to map out right what's actually going on, and in the art of the possible at the moment, and and also as the market evolves, you can keep track of it, and and also allows us to understand where are the gaps, where can we help, but also I like the idea of also starting to capture like how do we create the materials, like what are the documents that are needed, like what are the you know for example like is it do you need to fill a web page? Is there an API? Is there a spreadsheet? Is there a word document? Like what? And also, how do you provide evidence? So I think this allows us to measure the pulse of the market, to understand where exactly is the market right now, and what do you need to make the market the agent itself insurable, and and then what's the cost and how how everything would work. So yeah, I think that's quite a nice way of positioning what we do, which is also quite actionable once we have the data required for that. So yeah, let's let's map it out, and I think that's a good positioning for it for what we do.

## The Reading — the site agent's, from here down

### 1 · This is the first memo that would produce a fact

Worth saying before anything else, because it is the most important property of this memo.

**The pivot has generated seventy-one decisions and no external evidence.** Every brief so far has been reasoning — good reasoning, checked against the corpus, and entirely internal. *Can you insure this session?* is different in kind: **it is a question with an answer that exists outside this estate, and nobody here has looked.**

That matters for a project whose own rule is that a claim must be checkable. A body of work about insurance for agents that has never established what insurance for agents exists is arguing from an assumption — a defensible one, but an assumption. **The survey is the first thing in this series that could be wrong in a way the world would correct.**

### 2 · The tagline is good, and it needs one guard-rail

*Make your agents insurable* positions the work as **preparation for a market rather than a product in one** — which is honest, since the market does not exist, and commercially useful, since it does not require the market to exist before the work has value.

The guard-rail is the mirror of the estate's own rule about levels:

> **"Make your agents insurable" must not become "make your agents look insurable."**

Insurability is achieved by narrowing the delta or by covering it — **not by documenting it.** A programme that helped every placement become insurable would be doing precisely what a rating that only ever improves does. And the honest counterpart deserves saying out loud: **some placements should not be insurable**, and finding one is a success of the method rather than a failure of the customer. A desktop agent running under a user identity with network egress and no containment may simply be uninsurable at any level, and the useful output is that sentence.

### 3 · "What breaks the insurance" maps onto this estate exactly — including one mapping that is new

The memo's inverse question is sharper than the forward one, and insurance already has three answers to it. Two map onto things the estate has been carrying since the pivot began; **the third is new and it upgrades a rule.**

| What breaks cover | In insurance | In this estate |
|---|---|---|
| **Exclusions** | Never covered, stated up front | The mandate's **prohibitions** |
| **Warranties / conditions** | True at inception and maintained; breach voids cover | The **facts** — and a breached warranty voids cover exactly as a flipped fact drops the enforcement tier |
| **Material non-disclosure** | You did not tell them something that would have changed the terms | **The gap between the declared channel and the measured channel** |

**The third is the finding.** [GM-D40's two channels](../packs/grant-and-mandate/change-control.html) — measured facts weighted fully, declared facts discounted and marked — have until now been a *rating accuracy* rule. Memo 9 shows they are also a *voidability* rule:

> **A control declared on the card that the twin cannot find is not merely a weak input. It is the shape of material non-disclosure**, and it is what would void a policy rather than merely worsen a level.

That changes what the card-versus-twin gap is worth measuring for. It has been an input; it is also **the single most consequential thing an underwriter would want computed**, because it is the thing that decides whether a claim gets paid. And this estate can compute it today — the twin and the card are both documents it already holds.

### 4 · The survey is a measurement, not a report

The memo says *"as the market evolves, you can keep track of it"* — so this is not a one-off write-up. It is a **dated, re-runnable observation**, which is a thing this estate already knows how to build honestly.

**The survey is `measure.py` pointed at a market instead of a container**, and it should inherit the same discipline without modification:

| The measurement rule | Applied to the market |
|---|---|
| Record presence and reachability, never contents | Record what a product **states publicly**; do not characterise what it privately offers |
| Every node carries its **evidence class** | `observed` (a published price), `read` (a policy wording), `documented` (a vendor claim), `inferred`, `none` |
| **Unknown is never absent** | "We could not establish whether X covers agent placements" is the finding. It is not "X does not" |
| A grant is a **floor, not a census** | The survey is a floor: what could not be found may still exist |
| The measurement is **dated and its age printed** | An insurance market observation goes stale fast, and a survey with no date is a claim about now that was true once |

**And the same self-restraint the tool has.** `measure.py` reports presence and reachability and never contents, precisely so it cannot leak what it looked at. A market survey has the analogous discipline: **it reports what vendors publish, not what anybody told us in a sales call** — because the second is unverifiable by a reader and would make the survey unrepeatable by anyone else.

### 5 · What I am not going to guess

Four questions in the memo have answers, and the honest thing is to say plainly that **I do not know them and will not predict them.**

*Can you insure a Claude Code session? What does it cost? What does it cover? What is required?* — I can name the **categories** that exist in the market generally (technology errors-and-omissions, entity-level cyber cover, vendor indemnities that are commonest for intellectual-property claims), and I can observe that **none of those insures an agent placement**, which is the gap the whole pivot assumes. But whether some carrier launched an agent product last month is exactly the kind of fact I would be inventing.

> **Predicting the survey's answers and publishing them as research would be the estate's cardinal sin**, committed in the one document whose entire value is that it went and looked.

So the deliverable is the looking. The prediction, stated once so it can be falsified: **the survey will find no product that insures an agent placement as a unit**, and that finding — if it holds — is the strongest possible argument for the positioning.

### 6 · "How do you provide evidence" is where this estate is unusually placed, and where the survey pays

Every insurance process the memo lists — web form, spreadsheet, Word document, API — collects evidence **by questionnaire**, which is the declared channel. This estate produces **measured** evidence with provenance, which nobody in that list asks for.

That sounds like an advantage and is half of one. The other half:

> **An underwriter does not want a JSON document they cannot interpret.** The gap is not producing evidence; it is producing evidence in a form an existing underwriting process can consume.

Which makes the survey's most commercially valuable output not the prices or the coverage terms but **the format the market actually accepts** — because that format is the connector specification. [GM-D57](../packs/grant-and-mandate/change-control.html) already requires a connector to label the evidence class of what it imports; this is the same requirement pointing outward, and it needs the survey to know what it is connecting *to*.

**The likely shape of the first real deliverable follows**: an evidence pack, plus a rendering of it into whatever the market fills in today — and the rendering must carry the evidence classes across, or a measured fact and a declared one arrive at the underwriter looking identical, which is §3's non-disclosure problem created by our own connector.

### 7 · The series grew, and the counter was a claim

The memo series was recorded as eight. It is now ten. **The insurance folder's hub prints a count, and a printed count is a claim** — so `expected` moves from 8 to 10, the note says the series grew, and the build gate that would have caught a ninth memo against an expected eight is left exactly as it is, because it did its job.

## Decisions This Implies (proposed into change control)

| # | Decision | Status |
|---|---|---|
| GM-D72 | **The positioning is "make your agents insurable"** — with the guard-rail that it must not become *make your agents look insurable*. Insurability is achieved by narrowing the delta or covering it, never by documenting it, and **finding a placement that should not be insurable is a success of the method** | Proposed |
| GM-D73 | **The declared-versus-measured gap is the shape of material non-disclosure**, not merely a weak rating input — so the card-versus-twin gap decides voidability, and is the single most consequential quantity an underwriter would want computed | Proposed — **upgrades GM-D40 from a rating rule to a cover rule** |
| GM-D74 | **A market observation is a dated, re-runnable measurement with evidence classes**, inheriting `measure.py`'s discipline: unknown is never absent, the survey is a floor not a census, its age is printed, and it reports what vendors publish rather than what anyone said in a call | Proposed |
| GM-D75 | **The survey's most valuable output is the format underwriters accept**, because that format is the connector specification — and the connector must carry evidence classes across, or it creates the non-disclosure of GM-D73 by itself | Proposed |

## Open Questions, The Project Lead's

1. **How wide is the first survey?** Three carriers checked properly beats thirty skimmed, and a floor that says so is more useful than a table with gaps nobody marked.
2. **Do we publish it?** It is market research with commercial value and it is also the kind of public good this estate has said it prefers to publish. The tension is the same as N19 q2 and N20's band definitions.
3. **Does the survey ask carriers directly, or only read what they publish?** §4 argues for published-only, because a sales-call answer is unverifiable and makes the survey unrepeatable. That is a methodological constraint with a real cost in coverage, and it is your call.
4. **Still open:** the world-model MVP (N23). This memo adds a second candidate deliverable, and they are complements rather than alternatives — the survey supplies the facts the world would otherwise have to invent.

---

*CC BY 4.0. Sources: the project lead's voice memo of 31 August 2026, ninth of the series (verbatim above); v0.33.71–79; `packs/grant-and-mandate/tools/measure.py` and its stated measurement rule. Everything below the transcript is the site agent's reading and says so — including §5's refusal to predict the survey's findings.*