Last week, Voyfai merged its first fully autonomous pull request — an agent noticed a latency regression in production, diagnosed the cause, wrote the fix, opened the PR, and handled the review feedback, all without a human starting the work.
How the Loop Finds, Tracks, and Fixes Problems Without a Human Nudge
Five stages chain together to go from production telemetry to a reviewed PR. The first agent reads Datadog continuously, looking for regressions, slow paths, and error patterns — not just matching on the word 'error' but detecting the shape of a problem. When it finds something actionable, it writes a Jira ticket with the finding, evidence, and enough context for the next stage.
A second agent picks up that ticket, clones the repository fresh, reproduces the problem, and implements a fix in code. It opens a pull request just like a human would. Then a third agent works through the review feedback — Copilot, Codex, and human comments. For mechanical comments, it commits the fix one change at a time. Then it stops, and a human approves and merges.
Concrete example: latency on an endpoint creeps up after a deploy. No pages yet, but anomaly detection flags the drift. The first agent writes a ticket: which endpoint, when it started, how big the change, and the deploy it lines up with. The second picks it up, reproduces it, finds a query change introduced by that deploy, and opens a PR that batches the call. Copilot flags a missing null check — the agent adds it. A human asks for a clearer variable name — the agent renames it. By the time an engineer opens that PR, the diff is small, the description explains the cause, and the review comments are already handled.
Why the Human Stays (and How We Measure When to Remove Them)
Voyfai’s team is not keeping the human in the loop because they believe a person must always have the final say. Two practical reasons: First, trust has to be earned with evidence, not granted because a demo looked good. Confidently wrong is exactly what these systems are good at the moment you stop watching. Second, and more interesting: every human approval or rejection is a data point. Does the human merge as-is, or push more commits? When they reject, was the diagnosis wrong, the fix wrong, or just not how they would have done it? How often does the loop open a PR that goes nowhere?
Those numbers tell the team whether the loop is solving real problems or producing plausible-looking PRs that waste everyone's time. Until that data gets boring — until the human approves almost everything almost unchanged — they do not know the loop is good. The reviewer is partly there to tell them when they are no longer needed. The human stays exactly until the evidence says they have earned the right to remove them, and not a day before.
The Real Shift: Moving Judgment Into Automation, Not Replacing Engineers
The direction is to automate the review itself and take engineers out of routine changes entirely. The loop already only touches well-scoped problems with a clear signal and a clear fix. Ambiguous ones — where the right answer might be 'change the product' rather than 'change the code' — start with a human and will stay that way. The goal is to leave on an engineer’s plate exactly the hard, ambiguous problems and the architectural choices about where the system should go. Not reading a diff that adds a column for the hundredth time.
Automating routine paths — find known class of problem, write obvious fix, clear mechanical comments — is not replacing engineers. It is moving them up the value chain. The best engineers were never valuable because they could type a fix for an N+1 query. They were valuable because they could tell which problems were worth solving and how the system should be shaped.
None of this is as new as it sounds. Tests run themselves, CI/CD pushes to production without anyone watching. What is different now is that older automation could only walk a path defined in advance. An agent is handed a situation with many possible paths — which file holds the bug, which fix is right, which checks to run — and it decides. The judgment moved into the automation. That is the whole story.
Next challenge: once machines open their own PRs and engineers are already opening far more than before, you produce more change than any human review process was built to handle. Taking the human out of routine review is no longer a someday idea — it became a problem Voyfai had to solve now.
Source: Nobody Asked This Agent to Open This Pull Request - It Did Anyway
Domain: hackernoon.com
Comments load interactively on the live page.