That “trust is the binding constraint” line is spot on. In student projects here we see it every sprint - the team that writes the cleanest code is usually the one that actually talks to each other before coding. When you coach a team that’s moving fast but losing trust, what’s the first signal you watch for?
Trust built from small verified steps applies well beyond code. A habit that helps is to state the size of a change before making it, for example "this touches two files and one owner, and I will confirm with her before merging." Naming the blast radius in one sentence lets reviewers set their effort to match. A simple test is whether a colleague could predict your next update from your last one.
A line-count limit helps with batch size, but I would also ask whether a reviewer can verify one behavioral claim independently. A small diff can still hide a large change in permissions or failure handling. Pairing each step with its intended behavior, evidence that would catch a regression, and a way to reverse it makes the trust-building work visible. Otherwise, splitting a large agent change into several small PRs can satisfy the limit without reducing the reasoning burden.
Batch size is also the easiest knob to enforce in tooling. A CI check that fails any agent PR over a few hundred changed lines forces the work into steps a reviewer will actually read.
That “trust is the binding constraint” line is spot on. In student projects here we see it every sprint - the team that writes the cleanest code is usually the one that actually talks to each other before coding. When you coach a team that’s moving fast but losing trust, what’s the first signal you watch for?
Trust is the primary currency of professional software development.
Too pricey for me but if I was still an employee somewhere I would go to bat for this.
Trust built from small verified steps applies well beyond code. A habit that helps is to state the size of a change before making it, for example "this touches two files and one owner, and I will confirm with her before merging." Naming the blast radius in one sentence lets reviewers set their effort to match. A simple test is whether a colleague could predict your next update from your last one.
A line-count limit helps with batch size, but I would also ask whether a reviewer can verify one behavioral claim independently. A small diff can still hide a large change in permissions or failure handling. Pairing each step with its intended behavior, evidence that would catch a regression, and a way to reverse it makes the trust-building work visible. Otherwise, splitting a large agent change into several small PRs can satisfy the limit without reducing the reasoning burden.
Batch size is also the easiest knob to enforce in tooling. A CI check that fails any agent PR over a few hundred changed lines forces the work into steps a reviewer will actually read.