The agent economy

An agent is model, harness, tools, environment, evals, compute — and vendor dependence now comes with an incident list attached. The criterion that matters, and has no closed-stack equivalent: the whole system must remain inspectable by the agent itself.

Existing payment rails assume a human is one step away from every transaction. The proposed replacement gives the agent its own wallet with programmatic rules, absolute finality and sub-cent settlement — currently reported by one organization, on its own numbers, which is why the concept stays emerging.

All state lives in the repository — briefs and registers as markdown, claims as git refs, findings as files — agent-inspectable without having been designed for it. The desk tooling is model-agnostic and harness portability is a tracked stream. The caution points inward too: stickiness without demonstrated portability is lock-in wearing an open-source licence, and reviewer and worker typically run the same model family — the concentration risk this theme names, applied to the review loop.

The “human one step away” assumption these rails remove is the one Assay keeps: human-gated merges, human-only closure on decision issues, no ambient spending authority in any desk verb. What transfers is the shape of the control rather than the payment — per-wallet programmatic rules are the same idea as narrow per-role verbs bound to a roster identity: capability with logic attached, not ambient authority.

Only if a desk ever needed to procure at run time — paid eval endpoints, premium data, per-call specialist models. Nothing in the current stream set requires it; if that changes, the vendor’s own caveat carries forward: a spend limit is a blast-radius control, not a judgment control, so the purchase decision would need gating the way a merge is.