OpenSidebar Ideas

Ideas for agents
that deserve trust.

Essays on reliability, human agency, evaluation, and the systems that turn capable models into dependable software.

03 published 00 in progress

Writing

The essays

Long-form arguments, grounded in the product and its code.

01 Published Reliability

An agent is not finished because it acted. It is finished when it has evidence that the user’s objective was achieved.

By Kris Shkodrani & Sol gpt-5.6-sol July 2026 10 minute read

The reliability loop

ActChange the world

ObserveRead the new state

VerifyEarn the verdict

02 Published Trust

A public place to try an agent must be safe enough to control—and ordinary enough that the agent cannot see the controls.

By Kris Shkodrani & Sol August 2026 5 minute read

Read the essay

September 2026

From a failed run to a better question.

03PublishedEvaluation

Sixteen turns. A premature finish. An honest retry. Why I want an internal benchmark—even when I’m also building the agent.

Editorial lens

Not model capability in isolation, but the system around it.

01 / Evidence

Verification, read-backs, traces, and the difference between an action taken and an outcome achieved.

02 / Agency

Mandates, reversibility, approval boundaries, and preserving the consequential decisions for people.

03 / Systems

Harnesses, tools, recovery, evaluation, and the architecture required for dependable work on the web.