Independent agent field notes / July 2026

AI agents fail
in the gaps.

Independent tests of model access, reasoning controls, orchestration, and the failures hidden between products.

Model accessReasoning effortAgent orchestrationProvider limits

Featured investigation

GPT-5.6 Sol / Hermes Agent

Why GPT-5.6 Sol Max and Ultra Are Missing in Hermes Agent

Hermes reaches the model, labels xhigh as Max, accepts a separate max value in its core, and still cannot recreate Ultra with one setting.

What this publication covers

Which model and mode your account can use, which provider route actually serves it, and where a third-party client drops the control.

Reasoning values, tool loops, subagent inheritance, context handling, and silent fallbacks.

Source snapshots, sanitized logs, official documentation, reproducible checks, and fixes that survive outside a demo.

Earlier evidence series

The original source-audit articles stay live for readers who need them. They no longer define the homepage or the publication's direction.