Claim Compilation Below Agent Experience¶

Status: positioning note Last updated: 2026-05-06 Scope: explains how Assay relates to the emerging Agent Experience (AX) layer without claiming to own AX, add a new receipt family, or add new Trust Basis semantics.

Assay compiles agent-system outcomes into bounded, scoped claims.

Those claims are useful for AX-ready systems, ad-hoc agent test runs, and CI review alike. The positioning is structural: if "Agent Experience" becomes a durable category, Assay is the evidence layer beneath it. If the term fades, Assay is still the claim-compilation layer for agent outcomes.

Context: AX Is A Layer Above Assay¶

A category called Agent Experience is forming around making products easier for agents to consume:

Netlify frames Agent Experience as the holistic experience AI agents have as users of a product or platform.
Stainless focuses on agent-ready APIs, generated MCP servers, typed SDK use, and documentation search for coding agents.
Nx describes agentic experience as increasingly important alongside developer experience, with CLIs and workflows shaped for AI agents.
OpenAI and Anthropic describe harness engineering and harness design as practical work around agent loops, scaffolding, long-running execution, and agent-legible environments.

The shared move is making systems easier for agents to consume.

Assay does not compete in that layer. Assay can operate around MCP and ships policy/evidence surfaces for agent tool use, but it is not a docs platform, SDK generator, MCP-generation service, hosted AX control plane, or eval platform.

This note explains the layer Assay does occupy.

What Assay Does¶

Assay performs claim compilation: it takes a bounded outcome or runtime signal and produces a portable evidence artifact plus a scoped claim about that artifact.

Every supported receipt path answers four questions:

What was observed? The bounded evidence body.
Where did it come from? Source artifact, digest, and reducer provenance.
What claim can be compiled? A stable Trust Basis claim id and boundary.
What is not claimed? Explicit non-claims such as correctness, safety, or completeness.

Trust Basis is Assay's machine-readable scoped claim artifact. It is not a trust score. It is a claim table keyed by stable claim.id values, with levels and boundaries derived from verified evidence.

The fourth question is the differentiator. Most evidence systems leave the claim boundary to the downstream reader. Assay tries to make the boundary part of the artifact: observed here, reduced this way, claimable only under this scope, and not evidence of broader truth.

Why Claim Compilation Is The Wedge¶

Adjacent tools produce useful outputs, but not scoped Assay-style claims:

Source	Their output	Assay compilation
Promptfoo	selected assertion result	bounded eval-outcome receipt and `external_eval_receipt_boundary_visible`; no model-correctness claim
OpenFeature	selected boolean `EvaluationDetails` outcome	bounded runtime-decision receipt and `external_decision_receipt_boundary_visible`; no flag-config correctness claim
CycloneDX	selected ML-BOM model component	bounded inventory/provenance receipt and `external_inventory_receipt_boundary_visible`; no BOM-completeness claim
Runtime trace events	observed capability events such as filesystem, network, process, or tool decisions	bounded runtime evidence and reviewable capability surface; no claim that policy is correct

In this model, the adjacent system produces the outcome. Assay compiles the bounded claim around that outcome.

The first three rows above are not hypothetical. They are the currently released claim-visible receipt families, with checked-in proof artifacts in Evidence Receipts in Action and source-of-truth metadata in the receipt family matrix.

Where Assay Sits¶

AX is about what agents experience when consuming systems: agent-facing contracts, SDKs, docs, MCP servers, CLIs, and harness ergonomics.

Assay is about whether the outcomes those systems produce are inspectable, portable, and reviewable.

AX layer
  agent-facing contracts, SDKs, docs, MCP servers, harness ergonomics
  examples: Netlify, Stainless, Nx, OpenAI/Anthropic harness work
        |
        v
Assay layer
  bounded receipts + Trust Basis scoped claims over agent outcomes
        |
        v
Evidence sources
  Promptfoo, OpenFeature, CycloneDX, runtime traces, CI test runs

Assay does not require AX. Many agent systems in 2026 are not AX-ready: they are plain tests around HTTP calls, local tools, function-calling interfaces, shell commands, or ad-hoc prompt loops. Assay can still review the capability evidence those runs produce.

AX-ready systems are one sharp wedge. The broader surface is every agent test run that produces capability evidence.

Assay Harness¶

Assay Harness sits next to Assay. It orchestrates recipes, gates, and reviewer projections over canonical Assay artifacts.

Harness can:

compose recipes that combine receipts from multiple sources;
run Trust Basis assertions and diffs across receipt sets;
project artifacts into CI summaries, JUnit, Markdown, and other review surfaces.

Harness does not:

run evals for Promptfoo;
serve feature flag decisions for OpenFeature;
produce CycloneDX inventories;
provide AX itself.

Harness coordinates claim artifacts. It does not take over the work of the upstream systems that produce source outcomes.

Non-Claims¶

To support AX-adjacent systems without competing with them, Assay does not claim to:

own Agent Experience as a category;
build a docs platform for agents;
generate agent-facing SDKs;
run evals;
replace runtime guardrails or platform egress controls;
certify correctness, safety, or legal compliance.

Stretching Assay into those claims would replace a clean adjacency with a contested overlap.

Operating Principles¶

Machine-readable first. Receipt schemas, receipt-family matrix entries, canonical diff JSON, and Trust Basis claim ids must be parseable by other tools without human translation.

Copyable proof. Released examples should link to small artifacts, versioned recipes, verified bundles, and canonical projections rather than floating narrative claims.

Bounded guarantees. Each receipt should say what it proves, what it does not prove, and where the boundary sits. Bounded evidence is more useful in review than broad assurance language that collapses under due diligence.

What This Note Does Not Commit To¶

No prediction that AX will become the lasting industry term.
No prediction about which AX-layer tools will dominate.
No claim that any Assay artifact satisfies a specific compliance certification.
No new receipt family, Trust Basis claim, Harness recipe, or public integration claim.

In One Sentence¶

Assay compiles agent-system outcomes into bounded, scoped claims, usable as evidence for AX-ready and ad-hoc agent systems alike, and gateable in CI through Assay Harness or Assay GitHub Actions.