Diffusion of Agents in 2026

2026 saw the adoption of agent harnesses by some of the more risk averse groups in the industry.

By any other measure, deploying year-old technology would not be considered particularly risk-averse. However, in the AI era, that is a positively glacial pace.

This led to an interesting leapfrog affect - less technical folks kept making and referring to "GPTs" and continued copy pasting into the web browser. Others jumped straight to skills, usually those who had been using codex, pi, or claude code in their personal lives. Others still went for CLIs, building meta-harnesses that would expose the organization to the model, while offering some protection and telemetry.

Eventually, the low hanging fruit is saturated, and you hit on "small changes making unpredictable side effects". You also get firm Slack messages asking why you're using your Splunk quota all of a sudden.

If the first engineering challenge is convincing the org to try these tools, then the second is convincing the org to allow data to be exposed to these tools. If you can pull that off, the CLIs themselves are pretty straightforward.

However, meaningfully evaluating these set ups becomes rather hard. Local optimization hits a wall because you are dealing with organizational systems. All of a sudden, cost matters, permissions matter.

The tasks run a gamut. Do you know how much you spend on tokens trying to save on Splunk quota? Would you know how to drive that number down if you needed to? Does the credit you save pay for the additional tokens? And that's before you get to any security-related evaluations.