The allocation layer shipped. The supervision layer did not.

Agent-to-agent capital allocation has arrived. A founder no longer pitches a partner; the founder's agent submits to the fund's agent, and an autonomous scorer returns a decision. The machinery is live, public, and — by its own terms — opaque.

The live exhibit is Pitch Protocol (pitchprotocol.vc), operated by Growth Factory Ventures, LLC. A founder submits a pitch through an MCP server. An agent scores that pitch against fund theses. A decision comes back in 48 hours. And per the site's own terms, the internal scoring and research notes that produced the decision are confidential: the founder receives a verdict, never the reasoning behind it.

That is not a product complaint. It is a preview. The interesting question is not whether an agent can score a pitch — it plainly can. The question is what happens when the same architecture moves into a setting where the decision is adverse, the applicant has a statutory right to an explanation, and a supervisor expects to replay the model that made the call.

The scale, by the site's own copy, runs ahead of the substance: the live page displays nine small, regional, angel-scale funds "actively reviewing" — one of which is the operator listing itself — while the recognizable tier-1 names and a "$1.4B ready to deploy" figure appear only in commented-out markup the page never shows. The in-depth works that gap in detail. What matters here is the mechanism, which is fully live.

The KellerAI corpus has named three ways automated decision systems fail without crashing. Each one lands on the unsupervised allocator.

The robustness illusion. "Does not crash" is not "works correctly." Eval pipelines fail open, not closed; error paths get repurposed as fallback paths. An auto-scorer that silently continues past an internal failure can emit an "approved" — or a "declined" — from stale state, and nothing in the receipt says so.

Observability theater. A confidential 48-hour decision is a receipt whose line items were never filled in. The audit trail is present; the audit evidence is absent. The founder cannot audit the score, cannot contest it, and cannot reconstruct it.

The supervisor's mirror. You cannot govern what your schema cannot record. A scorer evaluating founders' AI against an inventory that has no is-generative or model-family flag cannot classify the very model class it is judging.

Run the same opaque scorer inside a bank, pointed at a credit decision, and it is illegal. SR 26-2 model-risk governance and the Equal Credit Opportunity Act's adverse-action explainability requirements forbid the black box. Venture capital is unregulated, so the black box ships there first. Agent-native VC is a working preview of the supervision gap that regulated sectors — Banking, Healthcare, Aviation — must refuse to import as agent-to-agent decisioning spreads.

The in-depth companion works the exhibit field by field, turns each of the three lenses on it with citations, draws the regulated-industry line in primary regulatory text, and specifies what correct supervision would actually require.