An approval is a pause before an action, giving a person a chance to inspect what will happen. It can make an agent workflow easier to supervise, but it is not a guarantee that an action is safe, correct or reversible.
The word "approval" does a lot of work in that sentence. A pause is only useful if the person can tell what they are approving.
Before approving a proposed GitHub issue or comment, check the destination, the content and the intended effect. Does it expose sensitive information? Does it make a promise on behalf of the company? Is it going to the right repository? Open the underlying context if the summary does not provide enough information.
If a tool asks you to approve something and shows you a paragraph describing the action, it has made your job harder, not easier. The point of the pause is to put the thing itself in front of you. A summary is a compression of the thing, and compression is exactly where the dangerous omissions live.
| Criterion | Should be visibleBefore you approve | Often missingThe usual gap |
|---|---|---|
| Destination | The exact repository, channel or inbox, spelled out. | A generic label like 'GitHub' with the target hidden behind a dropdown. |
| Content | The full text, not a summary of it. | A one-line description with the real payload underneath. |
| Scope | What the action will and will not change. | That the action is 'safe', with no scope stated. |
| Reversibility | Whether it can be undone, and how. | Nothing — every action reads as equally permanent. |
| Credentials | Which account and permission the action will use. | That the action runs with a shared token and no scoping. |
A read-only snapshot and a write action have different consequences. Give a connection only the permissions it needs, and understand which actions are actually implemented. In the current Founder2 implementation, GitHub issue creation and comments use approval paths. Do not assume that all imaginable external actions are available or protected in the same way.
Where a provider supports it, prefer a credential that can only do the one thing. A token that can read issues but not create them is a better token than one that can do both, even if the second is what you would eventually want. You can always add permission; removing it later is a different job.
An approval protects against the action being wrong. It does not protect against the action being wrong in a way you did not think to look for.
The most common failure is not a malicious action. It is a plausible one aimed at the wrong target: an issue written for the public repository instead of the internal one, a comment that quotes an internal figure, a change that is technically correct but lands before the product decision it depends on. These are all readable, and none of them is flagged by a warning.
The second common failure is approval fatigue. When every action needs a click, people stop reading them, and the approval becomes a rubber stamp with a good reason attached. This is the reason to keep the approval boundary narrow: an approval that appears once a week, for something that matters, is worth something. An approval that appears twenty times a day for minor edits is a speed bump that teaches people to say yes without looking.
Record the reason for important decisions, the evidence available at the time, and the person who made the call. After an approved action, check the result in the destination tool. If it failed or had an unexpected effect, correct it there and update the plan.
A decision trail is worth more to a small company than to a large one, because a small company has no memory other than the one it writes down. The founder who leaves in six months is, in practice, a different founder — and the only thing that carries the reasoning across is the record.
AI can draft a useful proposal quickly; the founder remains responsible for deciding whether it fits the business. For the wider picture, read what Founder2 can and cannot do today.