跪拜 Guibai
← All articles
Frontend · Backend

Codex Didn't Get Slower — It Started Acting Like a Full Engineer

By 掘金者阿豪 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Agentic coding tools are shipping into production editors now, and the default behavior is shifting from "assist" to "own." Developers who don't adjust their prompting and session habits will pay a latency tax on every small change — turning a two-minute fix into a twenty-minute audit.

Summary

A routine image-reference fix that once finished in minutes now triggers a cascade of review, verification, and quality-check steps that can run for hours — even after the code change is already applied. The shift stems from Codex evolving from a prompt-driven code assistant into an agent that treats every request as a full engineering task: analyze boundaries, scan the project, split work, modify, verify, review, and re-verify.

Plan Mode no longer means "output a plan only"; it means "plan first, then execute." The underlying workflow has become agentic regardless of the mode toggle, so even small requests inherit heavyweight process overhead. Long-running sessions compound the problem by accumulating massive context that the agent feels obligated to reconcile.

The practical fix is deliberate scope management. Explicitly instructing Codex to avoid global scans, skip unrelated modules, and prioritize direct modification over exhaustive review brings response times back down. The larger lesson is that agent capability is not free: matching the workflow's depth to the task's complexity is now a skill developers must learn.

Takeaways
Codex now runs a full agent loop — plan, scan, modify, verify, review, re-verify — even for single-file changes.
Plan Mode no longer suppresses agent behavior; it only changes how the task begins, not whether the agent workflow executes.
Code changes often appear in the editor long before Codex considers the task complete, because file modification is just one step in a longer chain.
Long-lived sessions accumulate context that forces the agent to reconcile unrelated history, slowing down small, isolated requests.
Explicitly constraining scope — "don't scan the whole project, only modify files directly involved" — cuts task time dramatically.
Tasks fall into three tiers: quick fixes (limit scope), medium features (allow analysis), and large refactors (let the agent run full workflow).
Agent capability introduces a new metric: total task-completion time, not just code-generation speed.
Conclusions

Codex's Plan Mode now behaves like a mislabeled feature — users who select it expecting a read-only plan get a full agent execution instead, which erodes trust in the mode toggle.

The agent workflow adds genuine engineering rigor, but applying it indiscriminately to trivial changes mirrors the organizational dysfunction of over-processed teams: a one-line fix shouldn't require a design review.

Session-length management is becoming a first-class performance concern. The convenience of a single long-running conversation directly conflicts with the latency cost of ballooning context, and the tool offers no obvious mitigation.

The shift from "code generator" to "task owner" changes the success metric. A 15-minute agent run that delivers a verified, working result may beat a 5-second generation that still needs 30 minutes of manual fixing — but only if the task warrants the overhead.

Prompt engineering now includes a new dimension: telling the AI what NOT to do. Negative constraints — "don't scan," "don't refactor," "don't expand scope" — are becoming as important as positive instructions.

Concepts & terms
Agent workflow
A task-execution pattern where the AI autonomously plans, splits work, calls tools, modifies code, verifies results, and reviews its own output — treating a user request as a goal to own rather than a prompt to answer.
Plan Mode vs. Agent Mode
In Codex, Plan Mode was originally understood as "output a plan only," while Agent Mode meant "plan and execute." The distinction has blurred: Plan Mode now plans first but still executes, making it a softer variant of agent behavior rather than a read-only mode.
Session context accumulation
The buildup of historical conversation, read files, diffs, logs, and prior task state within a single chat session. Large context helps complex multi-step tasks but adds latency and distraction for small, isolated requests.
From the discussion
Featured comments
Gainax 1 likes

I thought my relay station was just too slow; turns out the slowdown is Codex's own fault.

掘金者阿豪

I now add a line at the end every time: don't use agent, just make the quick fix directly.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗