Agents follow the patterns in the surrounding code more reliably than they follow your rules, so every inconsistent naming scheme and half-migrated architecture is a lesson you're teaching them. The fix isn't another rule file. It's deleting the bad example.
Two agents implemented and reviewed a full architecture refactor across five iterations, with zero lines of production code written by hand. What neither agent caught was four innocent lines in a coordinator that had quietly started making policy decisions.
Codex got the same god class Claude had failed to remove, plus one thing Claude never had: a target architecture. An implementing agent and an independent reviewer then took three rounds to converge, because agents renovate around bad architecture rather than delete it.
Handing an AI your engineering conventions cleans up the code without fixing the architecture. A 5,000-line god class shrank to 500 lines across tidy domain files, and an independent reviewing agent still returned FAIL: authority never left the central object.
Hitting a wall with AI usually means hitting the wall of the chat box, not the wall of AI itself. Give a model memory, tool access and a browser, and the interfaces we built for humans become the integration. The bottleneck stops being the model and starts being your imagination.
Synthetic checks like curl and Lighthouse measure a request your users never make. A HAR file captures the real logged-in flow, and handed to Claude alongside the codebase it turns a wall of JSON into a ranked list of fixes that get implemented in the same conversation.
A back-end engineer set out to redesign a React app with an AI agent and came out actually knowing React. The unlock wasn't the prompting, it was understanding the code well enough to steer it. As the cost of entry collapses, knowing the technology matters more, not less.
Handing everyone an LLM subscription gives them a model. A shared, self-hosted agent gives them a configured environment with company context, connected tools, and clear boundaries. Why a company should own that layer rather than rent it, and how to start small.
Coding agents made implementation cheap, but validation never got cheaper. That shift moves the bottleneck from writing code to deciding what is worth building, and the response is more autonomy and faster iteration, not more detailed upfront plans.
A written style guide is a museum piece nobody enforces. An LLM changes that: it reads rules in plain prose and applies them on every pull request, reaching judgment calls no linter could express. The style guide stops being a document and becomes a runtime.
An autonomous email agent hit a missing script, spiralled through 25 pointless shell calls, then fabricated email content including a prompt injection, and acted on it. The fix is not more warnings. It is structural validation before the model ever sees the data.
Using Claude Code as the orchestrator and Codex as the engine gives you the best of both tools: Claude's UI, Codex's depth, and minimal Claude token spend. The economics only work when the task is big enough to justify two layers.
The bottleneck in software engineering has shifted from typing to understanding. As AI handles implementation, engineers who thrive will be those who recognise patterns, curate guardrails, and connect technical decisions to business outcomes.
Hermes is an open-source AI agent that runs on a server, remembers across sessions, and builds reusable skills over time. The shift it represents: AI moving from something you summon to something that runs.
A customer support AI agent built in stages: shadow mode first, internal notes second, auto-send only after the data earns it. A walkthrough of the architecture, the knowledge base design, and the lessons that held up.
