PulseFlow Tecnologia
← Insights
·6 min read·executive-brief·coding-agents·codex·github-copilot
PulseFlow Tecnologia

The rise of coding agents and the impact on software engineering

Three pieces of evidence, from three different companies, point to the same structural shift: coding agents stopped being individual productivity assistants and started operating inside the software delivery cycle itself - issues, branches, tests, pull requests - as participants in the process, not just tools supporting it.

The experiment that tested the limit

OpenAI documented a repository that started empty in August 2025 and reached roughly a million lines of code in five months, with more than 1,500 merged pull requests, produced by just three engineers driving Codex agents - no line written manually by a human. Codex itself runs inside an isolated cloud container, with no internet access, reading and editing files, running tests and linters, iterating until it passes. Inside OpenAI itself, the most common use isn't open-ended feature generation - it's exactly the kind of task that overloads human engineers: refactoring, renaming, writing tests.

The flow that already exists inside GitHub

While OpenAI tested the limit of autonomy, GitHub documented how the Copilot coding agent already operates within a workflow any engineering team recognizes: someone assigns an issue to Copilot, the agent opens a draft pull request on a dedicated branch, explores the repository, makes changes, runs tests and lint, updates the PR, and requests human review - with the team able to iterate by commenting @copilot directly. As GitHub itself sums up, "every step is logged, visible, and open to team participation."

Two speeds, one direction

What separates the two cases isn't the underlying technology - it's the degree of autonomy each organization decided to grant. Copilot coding agent operates within any team's existing review process; OpenAI's experiment went further, testing how far a repository designed specifically to be "legible" to agents can sustain production with almost no human intervention. Both cases, though, depend on the same ingredient: explicit context about architecture, conventions, and boundaries, available inside the repository itself - not in conversations or external documentation the agent never sees.

What changes in engineering's operating model

The real impact isn't measured in lines of code generated - it's measured in what kind of work now flows through a different path than human-only. Low-to-medium complexity issues, test coverage, refactoring, and documentation are natural candidates for this new path; architecture decisions and final review stay human. Teams that already have a mature PR process (mandatory tests, automated lint, structured review) absorb this new participant with less friction - because the agent enters the same flow that already existed, only who initiates the work changes.

Sources