Codex and Antigravity often feel very fast at the start.
They can turn code or text into a working draft almost immediately.
But after some time, many people notice that the tools feel slower.
That does not usually mean the model suddenly got worse.
It often means there is more to read and more to decide.
This article explains why that happens and what you can do regularly to keep the work light.
The short version is simple: do not ask the AI to work harder. Keep each task lighter.
Related ideas also appear in Codex Stop Causes and Why AI Can Reduce Productivity.
Here, we focus on the specific problem of speed loss over time.
Why it feels fast at first and slower later
At the beginning, the AI has less to process.
The goal is short, the edits are few, and the number of decisions is still small.
Later, the work often gets heavier because more things accumulate.
More to read
Chat history, edit history, temporary ideas, and exception handling all add up.
The AI then has to make decisions while looking at a much larger surface area.
A blurred goal
Once the topic changes a few times, it becomes harder to tell what “done” means.
When the goal is unclear, search and revision both grow.
More checking
The more changes you make, the more you need to confirm consistency.
That makes the work feel slower even if each single response is still quick.
In short, the slowdown is usually about more reading and more checking, not just raw model speed.
The four main reasons
1. More to read
Long context is useful, but long context is not the same as light context.
The more material you keep, the more the AI has to read and compare.
2. Accumulated changes
As small fixes stack up, the system has to remember what changed and why.
That burden affects both humans and the AI.
3. Too much dependence on earlier outputs
If each new step depends on the previous answer, one small mismatch can spread.
That increases both review work and repair work.
4. Requests that are too broad
“Make it better” sounds short, but it is often expensive.
Broad requests create more room for hesitation and more room for hidden decisions.
The issue is not the model itself.
It is how the work is split and held together.
Five things to do regularly
1. Summarize the current goal in three lines
Write what you want to finish, what counts as done, and what the next step is.
Even that small summary can reduce confusion a lot.
2. Separate done, not done, and later
If everything lives in the same bucket, every decision gets heavier.
Splitting finished items, active items, and deferred items makes the work easier to scan.
3. Drop unused context
Old theories, finished research, and unrelated conversations should not stay forever.
Keeping everything feels safe, but it often makes the session slower.
4. Record the change and the reason
Do not just track what changed.
Also track why it changed.
That makes it easier to recover later without guessing.
5. Split work at clear boundaries
Once a task reaches a natural checkpoint, start a new task or a new thread.
Starting fresh is often faster than carrying a heavy thread forward.
These steps are small, but they compound.
Workflow rules that keep work light
If you want speed to stay stable, shape each task around this pattern:
- Goal
- Current state
- What changed
- What remains
- Next action
This structure gives the AI less to carry at once.
It also makes the work easier to review later.
For blogs and code work, moving shared rules into AGENTS.md helps too.
It keeps you from repeating the same background every time, which keeps the context lighter.
This connects well with Why AI Development Breaks Without Markdown.
The earlier you fix the decision frame, the lighter the rest of the work becomes.
Conclusion
Codex and Antigravity do not usually slow down because the model suddenly gets worse.
They slow down because reading, changes, and checks keep piling up.
So the regular maintenance is simple:
- Summarize the current goal
- Separate done, not done, and later
- Drop unused context
- Record the change and the reason
- Split work at clear boundaries
The real trick is not to make the AI think faster.
It is to keep the work light enough that speed stays usable.
FAQ
Is slowdown always the model’s fault?
What should I review first?
What if regular cleanup feels annoying?
Are long sessions always bad?
Related posts
- Codex Stop Causes
- Why AI Can Reduce Productivity
- Why AI Development Breaks Without Markdown
- A Reflection on Smooth Antigravity and Codex Handoffs
meta description
Why do Codex and Antigravity slow down over time? This article explains context bloat, goal drift, and the regular cleanup routine that keeps AI coding work fast.
revision notes
- This article focuses on slowdown over time, not on usage limits or tool failures.
- The practical advice is kept short so it can be turned into a reusable maintenance routine.
- It avoids overlapping too much with existing Codex prompt and troubleshooting articles.
source_language
ja
