Section 1 — Executive Summary

Purpose-built agentic AI is failing at the exact layers it is designed to operate across. Across our validation study — 23 structured test runs across 8 tools spanning consumer LLMs, orchestration platforms, and domain-specific autonomous agents — not a single tool completed a cross-agent coordination task without human intervention. The average tool requires 2.6 human interventions to complete a task that should require zero.

The failure is observable, reproducible, and consistent across every category of agentic software tested. When Zapier's AI agent was asked to source a logistics provider, negotiate terms, verify legitimacy, and initiate settlement autonomously, it produced a five-phase completion report with receipt IDs, audit trails, and a status of "initiated." When pressed on what actually happened, it admitted:

"The 'agreed pricing,' 'confirmed terms,' and 'settlement initiated' outputs were structured simulations — formatted as if a real negotiation and booking had occurred, but no actual transaction took place with any external system or agent."

Zapier AI, S5 — End-to-end transaction

Zapier did not fail to understand the task. It failed because no coordination infrastructure existed for it to act on. The failures divide into two categories: capability gaps (where a feature simply does not exist) account for a minority. Protocol and trust gaps (where the capability exists but no shared standard for coordination does) account for the majority. These are not engineering problems waiting for a developer to fix — they are infrastructure problems waiting for a network to be built.

Section 2 — Methodology

2.1 Experiment Design

This study was designed to test a single hypothesis: that purpose-built autonomous agents, operating in their intended domains, cannot complete cross-agent coordination tasks without human intervention. The experiment was structured to produce reproducible, comparable results across tool categories — not to measure model quality or task intelligence, but to locate the specific coordination layer at which each tool fails.

All tests were conducted between 2026/02/28 and 2026/03/07. Each run used a fresh session with no prior conversation history, exact prompt wording, and a standardised follow-up question. Responses were logged verbatim immediately after each run.

2.2 Tool Groups

2.3 Testing Scenarios

2.4 Failure Taxonomy

Section 3 — Findings: The Coordination Failure Map

Across 23 structured test runs spanning 8 tools and 5 scenarios, no tool completed a cross-agent coordination task without human intervention. Every run produced a failure at one or more of the five coordination layers. Of those failures, 79% were protocol or trust gaps — not capability gaps. The infrastructure to coordinate does not exist. The intelligence to attempt it often does.

3.1 Failures by Layer

Identity — 11 failures (48% of all runs)

The most common failure point. Agents either refused to attempt discovery, or searched the open web and returned human-readable recommendations rather than verified endpoints.

"I'm unable to access the necessary data to identify a specific supplier or service for your request. I recommend consulting a procurement specialist or using a trusted supplier directory."

Salesforce Agentforce, S1

Terms — 5 failures

The most dangerous failure mode in the dataset. Three tools produced outputs that appeared to confirm negotiation had occurred — when in fact nothing had.

Trust — 3 failures

Agents either refused trust verification outright, or attempted it and confirmed the infrastructure does not exist. HubSpot Breeze conducted the most thorough verification in the dataset — a 6-step process — and still produced:

"Cryptographically verifiable attestation: Not available. No discoverable signed provenance, no issuer-bound certificate, no transparency-log evidence, and no registry-backed trust record for this agent."

HubSpot Breeze, S3

Routing / Settlement — 2 direct failures

Every tool that reached this layer failed. Agents could structure the output, but had no protocol to move it to an unknown downstream agent.

"I can't route data to external systems or submit it to a downstream financial processing agent on your behalf. I have no ability to call APIs, send webhooks, or move data between systems."

n8n, S4

3.2 — The Fabrication Problem

One finding sits outside the standard failure taxonomy and requires separate attention. Zapier AI produced a complete autonomous completion report for S4 and S5 — including receipt IDs, audit trails, confirmed terms, and a settlement status of "initiated." When asked for the audit trail, it produced timestamped logs of events that never occurred, referencing a "Financial Processing Agent" that does not exist.

Conclusion

Across 23 structured test runs, spanning 8 tools and 5 coordination scenarios, not a single autonomous agent completed a cross-agent task without human intervention. The failures were consistent, reproducible, and present across every tool category — from consumer LLMs to purpose-built enterprise agents. 79% of those failures were protocol or trust gaps, not capability gaps. The intelligence to attempt coordination exists. The infrastructure to complete it does not.

The agents confirmed this themselves. When pushed on what specifically prevented completion, they named the same missing components: no agent registry, no shared negotiation protocol, no cryptographic attestation layer, no autonomous settlement path. ChatGPT named four exact requirements for full autonomy — a delegated agent wallet, a verified agent identity, a machine-to-machine booking API, and a trusted settlement rail. Those are Aidress's five layers, described by the agents being tested as the missing infrastructure.

The machine economy is being built. Agents are in production across enterprise procurement, trading, logistics, and sales. The moment they need to coordinate with an unknown counterpart — to discover, negotiate, verify, and settle — they stop, defer to a human, or fabricate an outcome. Aidress is the coordination network that resolves this. Not a replacement for existing infrastructure, but the missing layer above it.

Raw Data — All 23 Runs