Pip: Wulf Kaal has spent two decades arguing that institutions matter — and now the institutions need to work for readers that don't have eyes.
Mara: That's the territory this episode covers. Kaal's recent posts map what autonomous agents actually select for, what happens when those agents try to improve themselves, and why trust and attribution are becoming the scarce goods of the machine economy.
Pip: Let's start with what agents want — and why your landing page is the least of your problems.
Agentic web and machine readers
Pip: The premise of What Agents Want is that the web now has a second readership — one that doesn't browse, doesn't scroll, and arrives with a task and a set of disqualifying questions. The real question is: what are those questions, and what happens if you fail one?
Mara: The post frames it as four sequential gates. Here's the spine of the argument: "Failure at any gate removes a service from the candidate set before quality is ever evaluated. The sequencing matters because agent traffic concentrates: one orchestrator's routing choice fans out across every worker it spawns."
Pip: So you don't lose the comparison — you never enter it. That's a different kind of invisibility than bad SEO.
Mara: The four gates are discoverability, parsability, trust, and transactability. On discoverability, the post notes that as of mid-2026, a single well-known catalog file can announce a site's MCP server, agent interface, and API from one index. On parsability, context is the scarce resource — structure beats prose, and anti-bot friction isn't a cost, it's a disqualifier.
Pip: And the trust gate is where the stack quietly admits it hasn't solved anything. Registries attest publication, not behavior — which is a polite way of saying the verified-publisher badge tells an agent who shipped the tool, not whether that counterparty has ever behaved.
Mara: Right. The post puts it plainly: "The badge is a testimonial. Agents need a prior." Three of the four gates have native machine primitives. Behavior has none.
Pip: Which is where reputation enters — not as a nice-to-have, but as the missing institution. The post argues that a reputation system converts history into a priced prior: machine-readable, incentive-compatible, identity-binding, and compounding across the network.
Mara: And Attribution Is Infrastructure picks up the layer beneath that. Before any agent can stake anything on a claim about a work, it needs cheap, deterministic answers to three questions: what exactly is this object, who made it, and where does the canonical record live.
Pip: It's SEO, then citation hygiene, then institutional design — turtles all the way down, except the bottom turtle is load-bearing.
Mara: The attribution post is direct about the stakes: "An agent that cannot resolve an identifier will guess, and a guessing agent is a misinformation engine with excellent grammar." The fix is unglamorous — stable identifiers, full author strings, content hashes, canonical records — but the post frames it as plumbing, and notes that institutions are plumbing that held.
Pip: So the reputation layer and the attribution layer are both infrastructure answers to the same structural gap — the agentic web has prices and payment rails but no institutions. That tension carries straight into what happens when agents start improving themselves.
Self-improving agent systems
Pip: Agent Graphs: Self-Improvement Is Not a Search Problem takes the engineering pattern everyone is celebrating and asks the uncomfortable question — what exactly are these systems improving toward?
Mara: The post identifies a hidden dependency in every self-improving system: an external scorer that cannot be argued with. The key line is this: "Remove the external verifier and the apparatus loses its anchor. What remains is a system optimizing against its own estimate of its own quality. That is not self-improvement. It is self-congratulation with version control."
Pip: The verifier, it turns out, is not a piece of software. It's an institution — and the post argues the field has been spending that institution down ever since it borrowed it from benchmark culture.
Mara: The proposed alternative is to treat a proposed self-modification as work: submitted, attributable, and validated by parties who hold something at risk in the judgment they render. The post is careful to note this doesn't claim validation markets discover truth — only that divergence from ground truth, where ground truth exists, can be measured and bounded.
Pip: And whoever writes the reward writes the species. Which makes the trust question less abstract than it sounds.
Trust and attribution infrastructure
Pip: AGI Makes Intelligence Cheap. It Makes Trust Priceless. pulls back to the widest frame — what a general intelligence does to a research program built on trust without trusted parties.
Mara: The post argues AGI's effect turns on two variables: whether capability concentrates or disperses, and whether verifying an output stays harder than generating one. On the second, the post is pointed: "More capable agents are also more capable colluders. Consensus gets easier to manufacture precisely as capability rises."
Pip: So the stronger the agents get, the more load-bearing the reputation infrastructure becomes — not despite their capability, but because of it.
Mara: The post closes on scarcity: when intelligence approaches zero marginal cost, what remains scarce is verification, provenance, and the willingness to stake a name on a claim. Attribution Is Infrastructure, from the first segment, is essentially the operational version of that same argument.
Pip: The bet was never that machines stay weak. It was that the harder they get, the more the old institutional questions bind.
Mara: Discovery, validation, attribution, trust — these posts are all pointing at the same gap: the agentic web has the rails but not the institutions.
Pip: And the institutions are the interesting part. Next time, we'll see whether the empirical results start filling that gap — or just make it more precisely measurable.
Wulf Kaal 