The Agent Card Is the New Control Document

Koko: Nine in ten enterprises are already using AI in at least one function. And agents actually running in production? Single digits. That is not a technology problem.

Sam: It is a paperwork problem. Which is either deeply embarrassing or deeply clarifying, depending on your mood.

Koko: It is clarifying, actually. Because the piece we are unpacking today makes a very specific claim about what that missing paperwork is and why its absence is the thing killing production deployments.

Sam: And the claim is not subtle. The argument is that a production agent with no control document is, quote, an anonymous actor holding production credentials. That is the default state of nearly every agent running in an enterprise today.

Koko: Anonymous actor. Holding production credentials. Let that sit for a second.

Sam: I have sat with it. It is uncomfortable. Because the moment you say it that way, you realize the risk is not hypothetical — it is structural.

Koko: So the piece — and the full title is The Agent Card Is the New Control Document — the piece is making a very clean argument: the governance gap between demo and production has a specific shape, and it has a specific fix.

Sam: The shape being: you can prove the agent works once, in a demo. You cannot say who owns it, what it is allowed to touch, when a human has to step in, what evidence it leaves, or what it is actually worth. Those five questions have no home.

Koko: And the fix is giving them a home — a single, versioned, configuration file attached to every agent. The agent card.

Sam: Which sounds obvious until you realize nobody is doing it. The piece reaches back to two earlier inventions to explain why this matters: the org chart and the bill of materials.

Koko: Right. Before the org chart, large organizations had people doing things. After it, every role had an owner, a scope, a reporting line. It did not make people more capable. It made who-answers-for-what visible.

Sam: And the bill of materials did the same thing for physical parts. A component on an assembly line is trusted not because someone remembers building it, but because it traces to a supplier, a specification, a tolerance band, and a test result.

Koko: The data sheet is the contract. This part does this, within these limits, verified this way.

Sam: And the argument is that autonomous software work has neither equivalent. We are in the pre-org-chart, pre-bill-of-materials state. Capability without legible accountability.

Koko: The analogy is exact, the piece says. And I think it earns that claim.

Koko: Let's get into the actual structure. Because the piece does not just say you need a card — it specifies every field, what it records, and what failure mode it closes.

Sam: And this is where it gets genuinely useful. There are thirteen fields. Owner, purpose, persona, process, provider and model, tools, data products, risk tier, human-approval triggers, outputs, eval pack, control pack, and value card.

Koko: Each one is described as load-bearing. Each one closes a specific failure mode that is, quote, where a pilot dies on the way to production.

Sam: The one that jumped out at me is the tools field. Because it is not just listing what the agent can do — it explicitly includes what it cannot do. That boundary is the, quote, blast radius. The hard limit of what the agent can do to production.

Koko: And the piece walks through a real example. A Month-End Closer agent operating in record-to-report. The manifest is right there in the piece — you can read every line of it.

Sam: And what it shows is how granular the control is. The agent can draft a journal entry. It cannot post one without human sign-off. That is a human-approval trigger. And a material adjustment — anything above a threshold — escalates automatically.

Koko: The piece puts it beautifully: those two triggers are, quote, the entire line between an agent that drafts the close and one that commits it to the ledger. And in a tribal-knowledge world that line is a comment in a script that no auditor will ever see.

Sam: A comment in a script. That is doing a lot of work as a governance framework.

Koko: Then there are three fields that are themselves full artifacts — not adjectives. The eval pack bundles golden cases, policy tests, grounding tests, adversarial cases, and a regression suite. The control pack bundles approvals, segregation-of-duties rules, evidence capture, restricted-action lists, and the audit trail.

Sam: And the value card bundles the baseline, the target, the value lever, the owner, and the attribution method. In the Month-End Closer example, the value levers are close days — targeting a reduction of two days — and audit-adjustment rate, with named attribution methods for both.

Koko: Named attribution methods. That is the part that should make CFOs sit up. Because right now most agents are measured on usage statistics.

Sam: Usage statistics are the vanity metric of the agentic era. The value card is the operating metric.

Koko: Here is where I want to push on the piece a little, because there is a real tension it surfaces but does not fully resolve.

Sam: Go.

Koko: The piece says the card is not the hardest thing to build. It is not even the smartest part of the stack. It is the part that lets the CFO, the CIO, and the audit lead all say yes to the same agent.

Sam: Which is true. But if it is not hard to build, why is essentially no one doing it? The adoption numbers in the piece — nine in ten organizations using AI, agents in production in the single digits — that gap has existed for a while.

Koko: The piece's answer is that enterprises solved the wrong problem. They bought the thing that proves the agent can do the work once. They skipped the thing that proves the enterprise will let it do the work ten thousand times.

Sam: And I think that is right, but there is a harder version of the objection. Creating thirteen load-bearing fields, three nested artifact bundles, independent versioning on seven components — prompt, skill, model, tool, data product, policy, eval pack — that is actually a significant operational lift.

Koko: It is. And the piece does not pretend otherwise. What it argues is that the alternative is worse. The alternative is silent drift — a prompt edited in production with no record, a model swapped with no version bump, a control environment broken by a change nobody logged.

Sam: The piece uses a striking phrase for that: the quiet deploy. And the lifecycle section makes the point that the card makes every transition a reviewable event rather than a quiet deploy.

Koko: The lifecycle has six states — draft, demo, beta, enterprise pilot, production, retired. And the gates between them are explicit. An agent cannot enter enterprise pilot without a baselined value card. Cannot enter production without a passing eval pack and a live control pack.

Sam: So the lift is real, but the preconditions are checkable instead of a matter of judgment in a deploy meeting. Which is actually a lighter cognitive load, once the system is in place.

Koko: That is the reframe. The card does not add bureaucracy — it replaces a much messier, more fragile informal process with something that is just written down.

Sam: Configuration versus tribal knowledge. The piece is very clear that tribal knowledge does not survive a reorg, a resignation, or an audit request. Configuration does.

Koko: So if you are a CIO or a CFO sitting on a portfolio of pilots that are not making it to production — what does this piece tell you to do?

Sam: First: diagnose honestly. The piece says most enterprises can prove an agent works in a demo and cannot answer the five governance questions — owner, permissions, human gates, evidence, value claim. Go ask your team those five questions about your top three pilots. The answers will tell you where you are.

Koko: Second: make the card the gate. Not the model. Not the prompt quality. The piece is explicit that the frontier model is a commodity — every competitor can clear the capability bar on the same terms. The differentiator is the harness, and the card is where the harness's decisions about a given agent are recorded.

Sam: Third — and this is the one I find most useful for finance leaders specifically — attach a value card to every agent before you fund it at scale. The piece makes the dependency explicit: no agent card, no value card; no value card, no scale funding. That single rule converts a pile of experiments into a managed portfolio.

Koko: And the value card is not a vague ROI claim. It is a baseline, a target, a value lever, a named owner, and a named attribution method. The Month-End Closer example has close days as a lever with pre-and-post cycle time as the attribution method. That is auditable.

Sam: There is also a forward-looking signal in the piece worth flagging. It predicts that uncarded agent in production will become an audit finding. Same category as an unsegregated duty or an unreconciled account.

Koko: Which means if you are not building toward a card standard now, you are building toward a compliance remediation later.

Sam: The piece also predicts a common schema emerging — owner, purpose, tools, data products, risk tier, approvals, eval pack, control pack, value card — portable across the harness and legible to a reviewer who never met the builder. So the architecture you build now is probably close to the architecture that standardizes.

Koko: Get ahead of the standard rather than retrofitting to it. That is a fairly actionable read.

Sam: What I keep coming back to is the closing line of the piece. The org chart made people governable. The bill of materials made parts governable. The agent card makes autonomous work governable.

Koko: And once it does, the conversation moves from can this agent run to what is this agent worth. Which the piece says is the subject of the next essay.

Sam: That is the right forward question. Because right now the debate is almost entirely about capability. Can it do the thing? The piece is arguing that the next era of the debate is value. What does it return, measured how, owned by whom.

Koko: Which is a much more familiar conversation for every CFO who has ever managed a capital allocation process. It is just a new asset class.

Sam: An asset class where the data sheet is the agent card, and you would not fund a part without a data sheet.

Koko: The Agent Card Is the New Control Document is on Koko Knows — the full piece has the complete manifest, every field definition, the lifecycle table, and all the sourcing. Worth the read in full if today's conversation landed for you.

Sam: And if your honest answer to the five governance questions is I do not know — that is the piece to send your team on Monday.