Operations | Monitoring | ITSM | DevOps | Cloud

Previously, I described some core approaches to validating agent written code: feedforward and feedback techniques. Feedforward techniques are about avoiding errors up front, for example by coming up with better prompts and planning strategies. Feedback gives agents a signal that they have actually achieved a task. Feedback is a key part of common agentic patterns like Ralph loops or the /goal commands in Codex and Claude Code: keep working until some known condition passes.

How to set up blue-green deployments with CircleCI.

This post covers some techniques to get agents to validate their work. We’ll cover why this is important, a variety of types of checks you can employ, and mechanisms you can use to enforce them.

AI agents can write code faster than any developer. But for most teams, the feedback loop hasn’t kept pace. The agent generates code, pushes it to CI, and minutes later a full pipeline run catches a simple linting error or a failing unit test. By then the agent has moved on. Getting back to a working state means rebuilding context from scratch and burning tokens just to fix something that should never have shipped in the first place.

AI coding agents produce code faster than most teams can validate it. Without a validation step between the agent and CI, every problem gets caught after the push, and feedback arrives long after the agent has lost context. Agents need consistent feedback while they’re working so that small failures get fixed locally and CI stays focused on moving code into production.

Codex is OpenAI’s coding agent, powered by the GPT-5 family of models. It reads your files, proposes edits, and runs commands directly in your local environment. It ships as both a desktop app and an open source CLI, and it extends through plugins that connect it to external tools and services. Like any AI coding tool, Codex is strongest when the code it generates gets validated automatically.

Local development and remote validation were always meant to work together: developers iterate on their machine, run a few manual checks, then push to CI to clear code for production. But AI development broke that balance, flooding CI with a volume of commits no developer has read, let alone tested. Chunk sidecars restore the balance: lightweight, preconfigured environments that run alongside your local workflow and validate changes as they happen.

AI coding agents are fast. They generate functions, refactor modules, and wire up boilerplate faster than any human. What they don’t do by default is enforce the conventions a specific team has agreed on: the lint rules, the review patterns that senior engineers flag on every PR. A generated diff looks clean until someone runs CI or reads it carefully.

When a job fails, gets canceled, or never runs, the work that still needs to happen afterward (cleanup, notifications, teardown) has no clean way to trigger. There is no easy way to express “run this no matter what” in your pipeline config without duplicating jobs or adding fragile workaround branches. Terminal jobs change that.

A rolling deployment updates running application instances in batches, replacing old instances with new ones while the application keeps serving traffic. The concept applies to any system that can run multiple instances of an application, but Kubernetes has it built in as the default deployment strategy. Kubernetes terminates an old pod only after its replacement passes the configured readiness check, so no requests land on an unready instance.