The boundary that matters most here is the third one — expected changes proceed, unexpected ones stop. Most agent setups either gate everything (and grind to a halt) or gate nothing (and merge surprises). A month of running BRIDGE is the real signal though: did the unexpected-file stop ever fire on something legitimate and slow a real task down?
I would keep that stop even if it sometimes pauses a real task. The cost of a short halt is easier to live with than a change the agent was never meant to make. After each halt I add the legitimate path to the allowed list, so the next run can proceed without loosening the rule.
The halt-then-allowlist loop is the right shape — the rule never loosens, it just gets more precise, and each pause teaches the boundary something. A short halt on a real task is cheap insurance against a change nobody asked for. Does the allowlist ever get pruned, or does it only grow as the agent's legitimate paths get mapped out?
I would prune it as well. When the task that needed a path is finished, that path comes back out, so the allowed list only covers what the agent still needs. If something was removed too early, it just halts once and goes back in after a quick review.
That's the right trade — the cost of pruning too early is one halt and a quick review, which is cheap compared to an allowed list that only ever grows. The interesting failure would be a path that keeps getting pruned and re-added: at that point the review loop is telling you the task boundary itself is wrong, not the list. Have you hit a path that cycled like that?
I agree that a path which keeps getting pruned and added back is pointing at the task boundary rather than the list. At that point I would stop adjusting the list and split the task, so the path belongs to the one piece of work that needs it.
The 122 PRs with no diff are the number I keep coming back to. The boundaries held and nothing unauthorized was written. What leaked was work. Because inspect_app could only return status and steering files, the planner used execute_dev_task for audits, and Kiro finished by opening an empty PR or by writing the investigation into the repo as a text file.
If the only completion signal is a pull request, questions become pull requests. A read-only investigate path that returns a report, plus a hard "no diff, no PR" rule, would have prevented those 122 empty PRs, and probably some of the investigation-only ones, without loosening a single write boundary.
Product leader and systems analyst in fintech and edtech. Founder of AnalystCraft. Writing about requirements, acceptance criteria, agent evaluation, and Test-Driven Product Development.
The reviewed-head check and the separate “merged but sync failed” state are both valuable. One boundary I’m curious about is the target branch: if main changes after review while the PR head stays the same, the resulting merge can contain a combination that was never tested together. Does BRIDGE rerun the relevant checks against the prospective merge result or require a fresh review when the base moves? Keeping the reviewed head fixed solves one race; the base can still change the behavior being released.
Kento, a month of running it and writing down the incidents as well as the lessons is more than most of these posts give you. The fixed repo and branch on the server side are what keep the chat side from widening anything, so put the effort there. Blocking on an undeclared file in the diff feels like the least fussy of those guardrails to me. Weekend reading sorted.
Your Lesson 4 number stayed with me: a 32-second median from PR to merge means the AI review is the gate. In the pipeline I built at work we put people at both ends instead. The agent cannot push; the orchestrator pushes only after a person moves the card to Git Push, and a lead checks a production preview before anything merges. Once your three merge conditions are in, do you keep AI-only merges, or hand the final merge back to a person?
Senior Software Engineer focused on backend, databases, infrastructure, and AI coding agents. Sharing real systems, practical fixes, and reproducible results.
Location
Indonesia
Education
Mikroskil University: B.Sc. Information Systems, 2014
One boundary that sits under all four: the instruction files the agent reads are themselves a write surface, and a diff will not show what they hide.
A single line can carry a visible instruction, then 150 or more blank characters, then a second instruction. A reviewer scrolling the diff sees one ordinary line; the model reads two. It is worth linting: count runes rather than bytes so a U+3000 space counts as one blank, skip fenced code blocks and markdown tables, and warn on any run of 150 blanks that is followed by more text.
Does BRIDGE diff-review changes to AGENTS.md and .cursorrules, or are those treated as configuration?
We're a place where coders share, stay up-to-date and grow their careers.
The boundary that matters most here is the third one — expected changes proceed, unexpected ones stop. Most agent setups either gate everything (and grind to a halt) or gate nothing (and merge surprises). A month of running BRIDGE is the real signal though: did the unexpected-file stop ever fire on something legitimate and slow a real task down?
I would keep that stop even if it sometimes pauses a real task. The cost of a short halt is easier to live with than a change the agent was never meant to make. After each halt I add the legitimate path to the allowed list, so the next run can proceed without loosening the rule.
The halt-then-allowlist loop is the right shape — the rule never loosens, it just gets more precise, and each pause teaches the boundary something. A short halt on a real task is cheap insurance against a change nobody asked for. Does the allowlist ever get pruned, or does it only grow as the agent's legitimate paths get mapped out?
I would prune it as well. When the task that needed a path is finished, that path comes back out, so the allowed list only covers what the agent still needs. If something was removed too early, it just halts once and goes back in after a quick review.
That's the right trade — the cost of pruning too early is one halt and a quick review, which is cheap compared to an allowed list that only ever grows. The interesting failure would be a path that keeps getting pruned and re-added: at that point the review loop is telling you the task boundary itself is wrong, not the list. Have you hit a path that cycled like that?
I agree that a path which keeps getting pruned and added back is pointing at the task boundary rather than the list. At that point I would stop adjusting the list and split the task, so the path belongs to the one piece of work that needs it.
The 122 PRs with no diff are the number I keep coming back to. The boundaries held and nothing unauthorized was written. What leaked was work. Because
inspect_appcould only return status and steering files, the planner usedexecute_dev_taskfor audits, and Kiro finished by opening an empty PR or by writing the investigation into the repo as a text file.If the only completion signal is a pull request, questions become pull requests. A read-only investigate path that returns a report, plus a hard "no diff, no PR" rule, would have prevented those 122 empty PRs, and probably some of the investigation-only ones, without loosening a single write boundary.
The reviewed-head check and the separate “merged but sync failed” state are both valuable. One boundary I’m curious about is the target branch: if main changes after review while the PR head stays the same, the resulting merge can contain a combination that was never tested together. Does BRIDGE rerun the relevant checks against the prospective merge result or require a fresh review when the base moves? Keeping the reviewed head fixed solves one race; the base can still change the behavior being released.
Kento, a month of running it and writing down the incidents as well as the lessons is more than most of these posts give you. The fixed repo and branch on the server side are what keep the chat side from widening anything, so put the effort there. Blocking on an undeclared file in the diff feels like the least fussy of those guardrails to me. Weekend reading sorted.
Your Lesson 4 number stayed with me: a 32-second median from PR to merge means the AI review is the gate. In the pipeline I built at work we put people at both ends instead. The agent cannot push; the orchestrator pushes only after a person moves the card to Git Push, and a lead checks a production preview before anything merges. Once your three merge conditions are in, do you keep AI-only merges, or hand the final merge back to a person?
One boundary that sits under all four: the instruction files the agent reads are themselves a write surface, and a diff will not show what they hide.
A single line can carry a visible instruction, then 150 or more blank characters, then a second instruction. A reviewer scrolling the diff sees one ordinary line; the model reads two. It is worth linting: count runes rather than bytes so a U+3000 space counts as one blank, skip fenced code blocks and markdown tables, and warn on any run of 150 blanks that is followed by more text.
Does BRIDGE diff-review changes to AGENTS.md and .cursorrules, or are those treated as configuration?