Verify AI-generated code the way you would verify any change headed for production. Check it against the intended behavior and the existing architecture, run tests you did not accept on the author’s word, scan what the change adds, and have a qualified person who understands it approve it before it ships. Coding agents now open and modify pull requests, and AI systems review them, so the pull request is changing shape. It is not disappearing. Current official guidance from the UK Home Office and GitHub still requires human understanding and approval.
What “the end of the pull request” does and does not mean
The phrases “the end of the pull request” and “post-human era” are provocative framings, not established fact. Two things are observable: coding agents open and modify pull requests, and AI systems review them. Neither shows that human review has become unnecessary.
As an Amazon Associate I earn from qualifying purchases.
The most detailed recent measurement is a 2026 study by Selvanayagam and Ghaleb of AI-attributed pull requests and review events. Its figures should be read on their own terms:
- 248,641 pull requests in the study’s dataset received at least one AI-attributed review.
- Among AI-attributed review events, 45,269 were cross-product (author and reviewer attributed to different AI products) and 208,145 were same-product. These count review events, not pull requests, so they should not be added to the 248,641 figure.
- Cross-product AI-to-AI review occurred in roughly 1.6% of identified agent-authored pull requests. This is the study’s own estimate and depends on its dataset and attribution method.
- Cross-product review volume rose by more than two orders of magnitude from 2025-Q1 to 2025-Q3.
The study defines “closed-loop” narrowly: AI appears as both author and reviewer. That definition does not mean no human was involved in those pull requests, and it does not show that AI review matches qualified human review. What the data describe is a change in workflow, not the end of the pull request.
#1 Best Overall
The UK Home Office engineering standard, in its “Use AI” section, states:
“Teams will retain full accountability for all AI‑assisted code and outputs. AI tools cannot replace human judgement, understanding, ownership, or responsibility for decisions, designs, or changes made to systems.”
That standard expresses requirements for the Home Office’s own organizational context. It is a useful model, not universal law.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Rank #2
A public discussion asks the plain version of the question this guide answers: “How do you verify AI-generated code before deploying?” It is one anecdotal example, not a representative survey.
No authoritative, directly comparable figure for defect rates in AI-written versus human-written code is available from the sources cited here, so this guide does not offer one. Treat any such number as needing its method checked before you rely on it.
A six-step verification sequence
Run these steps in order. Early steps define what the later ones check against.
Rank #3
1. Establish the contract before reading the implementation
Turn the task into observable requirements and into the behaviors the change must never produce. Then compare the change with the ticket, design note, API contract, threat model, and architecture the project already uses. GitHub’s review guidance asks reviewers to check whether the code solves the right problem and follows project conventions. Ask these questions before reading the code:
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →- What did the agent assume about users, business rules, permissions, and failure handling?
- Which invariants must hold (for example, “a refund never exceeds the original charge”), and where are they enforced?
- What should happen on timeout, partial failure, or malformed input?
2. Read the complete change and confirm its provenance
Read the whole diff, not only the parts the summary highlights. That includes generated tests, configuration, dependency manifests, workflow files, and any changes to permissions or build settings. Confirm which agent and task produced the change and who requested it. GitHub documents Copilot-authored commits, co-author attribution, signatures, session logs, and audit events for its cloud agent, which together make that trail available.
3. Run independent functional and structural checks
Build or compile the change, treat every new warning as a finding until someone explains it, and run the existing test suite. Then add tests derived from the contract rather than from the code. NIST’s reviewed testing guidance (page updated October 6, 2026) describes three families that map directly onto this work:
- Black-box tests derived from requirements, covering invalid inputs, boundary values, and combinations of inputs.
- Structural tests based on the implementation, such as exercising the branches the new code adds.
- Regression tests built around previous bugs in the same area.
Tests written alongside the code by the same agent can share its assumptions. Generate boundary and invalid-input cases from the contract first, then compare them with the agent’s own tests.
4. Probe security and dependencies
Review each new dependency for existence, maintenance, provenance, license, and known vulnerabilities. AI tools can suggest packages that do not exist or that are suspicious look-alikes, so confirm the name, publisher, and source against the registry you actually install from. Then layer the security checks:
Free tools Windows power users keep installed
One-click scans. No signup required.
- Static analysis of the changed code, such as GitHub’s CodeQL where your repository uses it.
- Dependency advisory checks on manifests and lockfiles.
- Secret scanning across the full diff, including test fixtures and configuration.
- Fuzzing for parsers and other input-heavy components.
- A web-application scanner run against a running build for network-facing software, which NIST recommends as dynamic security testing.
- Continuing vulnerability monitoring of included libraries after merge, not only at review time.
5. Review AI-specific failure modes
Look for the patterns that AI-written code tends to show:
Best Value
- Calls to APIs, flags, or library functions that do not exist in the version you use.
- Contract constraints that were ignored or quietly reinterpreted.
- Logic that reads correctly but violates intended behavior in an edge case.
- Failing tests that were deleted, skipped, or weakened instead of fixed.
- Code that is plausible in isolation but inconsistent with project conventions or surrounding architecture.
- Maintainability problems that will cost the next developer time.
Ask any reviewer who raises a finding to explain why it matters and how to reproduce it. A finding without a reproduction is an open question to resolve, not a verdict. A second model can help surface candidate issues, which makes it useful during review.
6. Require accountable approval and a recovery path
The Home Office standard requires AI-assisted output to be reviewed and approved by suitably qualified people before production, requires teams to keep full accountability, and asks that AI-assisted changes be traceable. It also tells teams to plan for incorrect or insecure output and to keep ways to detect, mitigate, and recover from failures. On GitHub, the Copilot cloud agent’s documentation states: “Draft pull requests created by Copilot cloud agent must be reviewed and merged by a human.” That describes GitHub’s product, not every coding agent or repository platform. In practice:
- Record which agent or model produced the change and link it to its task.
- Confirm the rollback path works before merging a change on a critical path, for example by a tested revert or a feature flag.
- Name the approving person. Do not rely on a generic approval that any team member can give without understanding the change.
What each check can and cannot show
Each control answers a different question. The table sets out what each one establishes and what it leaves open, so that no single passing result is read as proof of correctness.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problems| Control | What it establishes | What it leaves open |
|---|---|---|
| Existing and contract-based functional tests | That the specified behavior holds for the cases run | Behavior outside the tested inputs and combinations. NIST treats tests as evidence about specified behavior, not as proof of all behavior. |
| Build and compile, with warnings reviewed | That the change builds in the target environment without unexplained warnings | Whether the built behavior matches intent |
| Static analysis | Matches for the patterns the analyzer knows | Vulnerabilities outside its rule set |
| Dependency advisory checks | Known vulnerabilities and maintenance problems in declared packages at scan time | Vulnerabilities published after the scan, and risks in code that is not declared as a dependency |
| Secret scanning | Credentials that match known patterns in the scanned content | Secrets in formats the scanner does not recognize |
| Fuzzing | Crashes or unexpected behavior from generated inputs to the targeted component | Parts of the input space the fuzzer never reaches |
| Web-application scanning (dynamic) | Exploitable behavior it can reach in a running, network-facing application | Routes, roles, and logic it does not exercise |
| Commit signatures, session logs, audit events | Which agent or person made the change, and when | Whether the code is safe or correct |
| AI reviewer | Additional findings and explanations that may catch issues | Independent assurance, unless it is shown to fail differently from the author and its output is validated. It is not equivalent to qualified human review. |
| Qualified human review | Whether the change fits intent, trade-offs, and ownership | Complete coverage. Quality depends on the reviewer’s expertise and the time allotted. |
Where platform controls fit
Hosting platforms are adding checks that run automatically on agent output. GitHub announced on June 9, 2026 that automatic security validation for third-party coding agents working in repositories is generally available, with CodeQL, dependency advisory checks, and secret scanning following each repository’s settings. These are GitHub-specific features and may change. The checks run only where the repository enables them, so confirm the settings before relying on them. Enabling them is worthwhile, but they do not replace steps 1, 5, and 6.
Quick Recap
Merge decisions
Turn the results into one of three outcomes:
- Block: a secret is exposed, a dependency cannot be verified to exist or be maintained, a deleted or weakened test is unexplained, or a security finding has been reproduced.
- Return to the agent: the change misses a contract requirement, lacks boundary or invalid-input tests, or rests on assumptions nobody has confirmed. Rerun with a corrected specification, and record the change rather than patching the output by hand without a trace.
- Approve with a named owner: the checks have run, every finding is fixed or formally accepted by the owner, and the approver can explain what the change does and how it fails.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
