Coding agents, context windows, and momentum on tap

Original draft (unedited)

Coding agents!

I’ve had lots of hands-on experience the last few months with Cursor, Codex CLI, across a few codebases. Generally a really positive and constantly improving experience so far. Even if it often misses the mark, or gets stuck in a loop after building a 750k context window, it just greases the wheels enough to keep momentum up on projects, making it easy to dip in and out of multiple sites and automations without the usual overhead of switching.

One thing I really liked was how you can even add feature requests from your phone in the evening and wake up to them ready for review. Or how you can set multiple agents off at once to build a page here, a script there, fix the UI here and build this feature there. Using both Cursor and Codex in the same repo too, if you wanted to review multiple approaches.

It does feel like it gives much more power than the earlier “auto complete” or chat UI copy-paste methods of last year.

I still have to investigate the use of sub-agents more, mainly because to-date I’ve doubted their use when the models tend to do what you’d ask sub-agents to do anyway (perfect the UI, test the code before delivering, sanity checks, look for relevant context etc.). But it will be worth testing still.

I did test local models, with Qwen Code and roocode, but didn’t get too far. It uses all system resources and operates at a lower level than the cheap frontier models. At least for now. With the new AI chips coming out, I suspect that situation will reverse in the next year. I may have an early upgrade situation on my hands! When that happens, a fully autonomous business doesn’t seem too far off the cards. An agent for sales, marketing, customer service, documentation, code fixes, finance etc. if the business is sufficiently simple.

The downside of projects getting too big too quickly where the mental model of the whole is not fully formed, leading to issues later, is still probably an issue in larger code-bases.

But for static site work, scripts and automations, fleshing out ideas for quick testing, it’s a real productivity multiplier. You can literally automate any business process very quickly nowadays, if you know your way around them. It’s like having a small team of coffee’d up devs on hand 24/7.

AI reworked version

Agents as momentum keepers

The last few months have been full of hands-on time with Cursor and the Codex CLI across a mix of codebases. The experience has been overwhelmingly positive, especially with how fast both tools are improving. Even when the model overshoots, stalls, or loops after chewing through a 750k token context window, it still greases the wheels enough to keep projects moving. Jumping between sites and automations no longer carries the usual mental tax; there is always an agent ready to take the next step.

Parallel requests, parallel progress

One of the standout perks has been how easy it is to queue work without being tied to a desk. Dropping in feature requests from a phone at night and waking up to implementation-ready diffs has become normal. Setting multiple agents loose at once—one building a landing page, another wiring up a script, a third polishing UI, a fourth shipping a feature—feels like running a compact, caffeinated dev squad. Using both Cursor and Codex in the same repository even makes room to compare approaches before choosing a direction.

Beyond autocomplete era thinking

Compared to last year’s autocomplete bars and copy-paste loops, the current crop of agents gives a real sense of leverage. They are no longer just smart suggestion engines; they are process partners that remember context, own tasks, and resurface edge cases before they land in production.

Still learning the sub-agent ropes

I have only scratched the surface with sub-agents. So far I have doubted their value because the base models already cover the obvious prompts—perfecting UI, running tests, sanity checks, context gathering. Still, it is a frontier worth testing, especially as orchestration becomes a differentiator.

Local models and looming hardware shifts

Local experiments with Qwen Code and roocode were less convincing. They devoured system resources and felt sluggish next to inexpensive frontier models. That might flip soon. With new AI-focused chips on the horizon, local-first workflows could reclaim the lead. If that happens, an early hardware upgrade could make a fully autonomous business stack—sales, marketing, support, documentation, fixes, finance—feel very close.

Mind the mental model gap

The main caveat is scale. When projects balloon faster than the mental model of the system, missteps accumulate inside the larger codebases. Momentum without architecture still creates debt. That risk has not gone away.

Where agents shine today

For static sites, scripts, automations, and quick idea validation, these agents are outright productivity multipliers. With the right prompts you can automate almost any business process in short order. It genuinely feels like having a round-the-clock bench of energised developers ready to ship.

Closing thought

Coding agents are not flawless, but they have already changed the cadence of work. The bottleneck is no longer “can I do this?” but “what should we build next?”