Codex keeps releasing better models and improving its harness to become more agentic. The latest updates seem to involve far more tool calling. More tool calls produce better results, but they also burn through my computer's local resources exceedingly fast.
It is now normal for me to have three long-running agents working in Codex at the same time. Each handles a separate workflow. I make decisions for the agents, correct course when needed, and then go do something else while they keep working.
This has become one of my most productive ways to work.
However, the latest Codex updates are crushing the CPU and memory utilization on my work laptop, which is a pretty good computer by any standard.
I know this is not a perfect technical comparison with what happens inside a data center. But the basic mechanism is similar. Agent harnesses need CPUs to run core computer functions. My agents spend much of their time using Python, working with files, and calling other tools. The model provides the intelligence, but the CPU still has to help carry out the work.
I never used Python before. I still do not know how to write it myself. Yet with Codex helping me, I am building useful software. Business logic and rules that once lived in spreadsheets, process documents, or my head can now become working code.
Long-running agents also need to preserve more state. Codex is good at automatically compacting its context, but I want the full history of a project to remain available in one place. I want the agent to remember the decisions we made at last week's steering committee meeting without forcing me to reconstruct everything.
That continuity takes a meaningful burden off the knowledge worker.
I will admit there are other ways around these resource constraints. Software will become more efficient. Some work can move to the cloud. I could also run fewer agents at once.
My response will be simpler: I am going to buy a more powerful computer so I can keep Codex hammering away at difficult problems.
The return is clear. I would need a team of multiple people to replicate the amount of development, research, and writing I can now manage through parallel agents. A new computer is easy to justify when it expands the amount of useful work I can supervise.
The bears can argue that CPU and memory bottlenecks are overblown. But models and agent harnesses are improving every week. Codex in August feels several times more powerful than Codex did in June.
The pace is head-spinning. If other knowledge workers adopt agents the way I have, agentic AI will not end the computer upgrade cycle. It may restart it.