A passing test suite is not a review
Tests check the cases someone thought to write. An agent that changes behaviour can leave every existing test green, because the test asserted the old behaviour only loosely, or because nothing covered the path at all. Green is evidence that the change runs; it is not evidence that the change is right.
The same applies to a tidy summary. The agent’s description of what it did is a claim, written by the same system that made the change. Read the diff, not the cover note, and treat the tests as one signal among several.
Start from a clean, reversible state
Before the first prompt: commit or stash your work, create a branch, and know exactly how to get back. A coding agent with terminal access can run migrations, rewrite configuration and delete files; a reversible starting point is what turns an accident into an inconvenience.
If the task touches anything regulated or confidential, use a scratch copy of the project and a conversation that contains no secrets. The provider’s terms and data-handling documentation decide what leaves your machine — read them rather than assuming the boundary.
Read the diff in the order it will hurt
A large generated change is easiest to review in a deliberate order rather than top to bottom.
- Entry points: routes, handlers, permissions and anything that changes who can reach what.
- Data: schema, migrations, serialisation and queries, especially anything that deletes or rewrites.
- Configuration: environment variables, build steps, CI and deployment settings.
- The rest: naming, structure, duplication and the cosmetic edits that make a diff look bigger than it is.
Then read the tests it added and ask what they do not cover. An agent that wrote both the code and the test can encode the same misunderstanding twice.
Dependencies, secrets and the bits nobody reads
New dependencies are the quietest risk in a generated patch: a package name that looks familiar, a version that came from nowhere, a lockfile rewritten wholesale. Check what was added, why, and whether the project already had something that does the job.
Scan for credentials, tokens and internal URLs that should not be in a commit, and check that logging does not now print customer data. These are the lines reviewers skim, and exactly the ones an agent has no reason to be careful about — it optimises for the request, not for your threat model.
When to throw it away
If the diff is larger than the task, if you cannot explain a hunk without guessing, or if the change solves the symptom rather than the cause, delete the branch. Starting again with a narrower request costs less than maintaining code nobody understands.
The same rule applies to your prompt: the review checklist in our Manus guide and the prompt-structure advice in our Prompt Cowboy guide both shorten the diff. For the product itself, start with our Freebuff overview, and how we handle referral links is set out in our affiliate disclosure.
Frequently asked questions
Should I let it run commands in my terminal?+
Terminal access is what makes a coding agent useful and what makes it risky, and the provider decides how permissions work. Start on a scratch copy or a throwaway branch, watch the first runs closely, and keep credentials out of reach; a command that only reads files is a very different request from one that deploys.
Is a free coding agent good enough for real work?+
It can be, for mechanical work you are able to review. Free is a commercial model, not a quality guarantee: the provider funds it with advertising and can change models, limits or features. The deciding factor is whether you can read and test what it produced, not the price.
Do I need to disclose that the code was agent-written?+
Follow your team, repository and licence rules, and answer honestly if you are asked. What matters legally and professionally is that a person reviewed the change and stands behind it; a generated patch that nobody read is the risk, not the tool that typed it.
Sources & further reading
Official provider documentation, reviewed 2026-10-26. Terms can change; verify the version for your country.