Coding agents have turned software delivery into a blink: a prompt vanishes into the IDE and, like a digital magician, a merge request appears five minutes later. That magic can be intoxicating, until the rabbit hops out holding a broken build, an unlinked issue, and a suspicious dependency. The real question is not whether an agent can type fast, but whether it knows where it is typing.
A friend once asked an agent to fix a sensor crash in a C++ repo. The agent happily patched the code, compiled it, and looked proud. Then CI erupted because the team had just added a linter rule nobody mentioned in the prompt. The fix was technically correct and practically useless, like bringing a clean shirt to a mud wrestling competition. Without lifecycle context, agents make shallow fixes and push them straight into review cycles.
I have also felt the context tax. Last month I wrote a careful prompt, watched the agent write code, and still had to explain the acceptance criteria, the review notes, and the pipeline config like a tour guide with a very forgetful robot. The moment I gave the agent access to the GitLab issue, the difference was like swapping a flashlight for the entire electrical grid. It read requirements, linked ‘Closes #32’, and stopped guessing.
The GitLab tutorials show three levels. First, repository-only: the agent sees files and your hopeful description, then invents a fix that compiles but may miss the team’s intent. Second, repository plus issue: the agent uses an MCP server to read acceptance criteria, milestones, labels, and notes, producing a change that fits the plan. Third, merge request: the agent receives review feedback, adds tests, updates docs, and commits again while CI validates it. Fewer review rounds. Shorter merge time. Same human gate.
For platform teams, context is governance. Issues list requirements. CI config sets the quality bar. Code owners define review expectations. Security scanners enforce policy. An IDE sees a file; the platform sees the lifecycle. That matters most when agents generate more code, because more code means more findings, more fix MRs, and more opportunities for false positives to waste a weekend. With full project context, security analysis filters noise and creates remediation merge requests without removing the human from the decision.
Another quiet hero is AGENTS.md. It tells the project what matters: structure, commands, standards, forbidden zones, and how to behave. In the Codex tutorial, it covered Rust edition, async patterns, and CI image pinning, so the prompt did not need to re-teach the universe. Yet context is not free. Models degrade when their windows are stuffed with irrelevant fragments. Platform teams should feed agents structured, relevant context, not a firehose of files, because a focused robot beats a confused genius.
The winning organizations will not just install agents. They will connect them to issues, pipelines, policies, and review. Speed without context is rework in a hoodie; speed with context is a team that ships safely while still going home at five.
