An AI assistant that drafts an email leaves you with text to review. Give it permission to send messages, update records or change files, and the job becomes more complicated. It has to understand the request, choose the right action and confirm what actually happened.
The appeal is less work for the user. But if every completed task needs to be reconstructed and checked from scratch, some of that work has simply moved. For anyone considering an AI agent, a practical question can be how much of a job they can hand over without having to do it again themselves.
Engineers at an AI company described a coding agent that confused file locations after moving between folders. Requiring tools to use each file’s full path resolved the issue in their tests.
Giving an agent a method
Consider software that suddenly starts running slowly. An agent might be able to read the code and retrieve performance measurements. It still needs a method for deciding which changes to examine, which clues matter and what would count as a useful fix.
Tommy Tran, a software engineer at Meta, worked on this problem in an agent platform for infrastructure efficiency. He told LeadDev that he combined finding software improvements and investigating performance problems within the same architecture.
Tran worked with senior efficiency engineers to capture their investigation methods. He recorded “the checks they run, the signals they trust, the order they investigate in,” he told LeadDev. These became reusable instructions, known as skills, that guide the agent through a particular kind of problem.
He separated those instructions from the tools used to inspect and act on the system. Specialists could then refine investigation methods while keeping the tools stable. That makes their expertise something the agent can consult during a task.
An Engineering at Meta article co-authored by Tran explains the division. Tools retrieve code, performance data and configuration history. Skills help the agent decide how to investigate. When a slow function handles converting data between formats, for instance, a skill can direct it to check recent changes to the data’s structure.
That can be useful beyond this particular setting. Access to a spreadsheet does not tell an assistant how a business defines revenue. Access to a customer record does not explain which circumstances justify a refund. Those decisions need instructions and evidence that the agent can use.
A proposed fix still needs a decision
Tran also set a boundary around the platform’s autonomy. It investigates and prepares a proposed fix, while a human engineer retains approval over changes to live systems, according to his LeadDev interview.
Meta’s account describes the preparation behind that handoff: the agent gathers relevant documentation, examples of earlier fixes and criteria for checking an optimization. When a software change causes a slowdown, the proposed fix goes to the engineer who wrote that change for review.
The company reports that automated diagnosis can reduce roughly 10 hours of manual investigation to about 30 minutes. That is a diagnosis figure; reviewing the proposed change and putting it into use are further steps.
Meta’s reported diagnosis time illustrates the intended benefit of this workflow: an engineer receives a proposed fix after the agent has completed the investigation. How much time that saves across the whole job still depends on whether the proposal holds up under review.
Human approval also takes time and attention. A reviewer needs enough context to assess the proposal, including what changed and how it could help. A system that produces suggestions faster than people can assess them could create a new backlog. The quality of the handoff is worth considering alongside the speed of the investigation.
AI engineers make a related distinction for coding agents: automated tests help check whether code works, while human review remains important for broader system requirements. A passed test and an approval answer different questions.
Measure the work that remains
For everyday users, the same issue appears in less technical forms. An email assistant should show whom it will contact and what it will send. An assistant editing a file should make its changes easy to inspect. One helpful question is whether the person can verify the outcome without having to repeat the entire task.
Costs need the same scrutiny. Developers at some AI companies are advised to start with the simplest approach that meets the need, noting that agents can trade additional waiting time and expense for better task performance. More autonomy is a choice to evaluate for a particular job.
A fuller comparison could account for the agent’s attempts, the time spent checking its work, any corrections, and the manual steps still needed. A fast first attempt may deliver little benefit if it leaves a person with a long repair job. A system with a narrower remit can still be worthwhile if it reliably completes the part it takes on.
That can give users a practical standard for judging the next agent they try. Look for clear permissions, visible results, and an understandable point at which the system asks for help. The benefit can be the time and effort saved after checking, and corrections are counted.
