When a production bug exposes a missing requirement, fixing the code is only part of the job. Update the spec and add a test tied to that requirement, so the next agent doesn't start with the same mistake.
馃 Made with AI
1
1
When an AI coding agent fails a check, send the failed condition back with the task. Without it, the next attempt has to guess what went wrong. A retry needs evidence it can act on.
馃 Made with AI
1
2
An engineer quietly rewriting an agent's patch is still rework. Keep that rescue attached to the original attempt, or your AI productivity report rewards the workflow for hiding its cleanup costs.
馃 Made with AI
2
2
7
If a cheap AI route saves a model call but creates rework, the routing score missed the real cost. Judge the decision by what happened next, and turn the miss into a test the next model has to pass.
馃 Made with AI
1
3
When someone fixes an AI routing mistake, keep the original decision tied to the correction. Otherwise the final success hides the failure, and tomorrow's run can need the same rescue.
馃 Made with AI
1
An AI judge saying "I'm not sure" only helps if the work has somewhere to go. Route the case to a stronger evaluator or a person, and keep the reason it stopped. A score in a log won't clear the queue.
馃 Made with AI
1
1
Your AI champion's repeated interventions are a map of what's missing. Each time they supply the same context or catch the same unsafe action, there's a rule or check the system still needs.
馃 Made with AI
1
1
Your editor can hide a missing requirement because you're there to fill it in. Before a coding task runs unattended, spell out what may change, what proves it worked and when the agent must stop and ask.
馃 Made with AI
1
1
Before an AI agent starts building, ask what evidence would make you abandon its approach. If the answer is missing, the research isn't finished. A working implementation won't settle that question.
馃 Made with AI
3
1
18
The model name can stay the same while your coding agent changes. Version the prompts, tools, context rules and retry policy, then run the same evaluations before promoting any change.
馃 Made with AI
1
2
2
Give narrow, reversible tasks a cheaper agent when independent checks can prove the result. Work that touches money, identity or stored data needs stronger verification. Price the whole route, including repairs.
馃 Made with AI
1
1
An agent can trigger a deployment without holding the deployment key. Every tool it calls needs to enforce the task's limits. Otherwise, handing off the work becomes a way to borrow permission.
馃 Made with AI
2
1
9
A bug fix should leave behind a way for the next agent to reproduce the failure and check the repair. If someone has to explain the same problem again next month, the fix left work unfinished.
馃 Made with AI
2
7
Telling an agent it can't release software means little if its credentials still allow it. Put the limit in the permissions, so it holds when the session or model changes.
馃 Made with AI
1
1
Correcting an AI agent's answer fixes one run. Fixing the missing check, stale context, or loose permission changes the next run too. That's how a delivery system learns without the model retraining itself.
馃 Made with AI
1
2
Before an agent retries a failed task, it needs to know what already happened. A fresh run can repeat an external action or erase useful failure state. Recovery starts with that record.
馃 Made with AI
1
1
An agent's retry policy decides how much authority survives a failure. Review what permits another attempt, what carries over from the last run and what forces a stop. That's part of your security boundary.
馃 Made with AI
1
4
Removing a feature means checking what it leaves behind: background jobs, stored data, permissions and callers. If the button is gone but its jobs still run, you haven't finished removing it.
馃 Made with AI
1
2



















