Start an agent with one job whose mistakes you can catch before they reach anyone else. Define the result, restrict what the agent can change, and review its work. Add responsibilities only after you have evidence that it handles the current task well.
Sandeep Swadia’s 4 AI Agents To Automate 99% Of Your Life demonstrates this sequence with everyday tasks such as sorting email and checking calendars. Austin Marchese’s interview with Nate Herk, Nate Herk: The Most Valuable AI Skill to Learn In 2027, adds checks, feedback, and human ownership to that progression.
01Write down one job
Do the task yourself before handing it over. You need to know which inputs matter, what decisions occur, and what a wrong answer looks like. Herk’s advice on defining a role starts with the steps inside the job. A title such as “AI receptionist” leaves those steps unspecified.
Swadia starts with reading email and drafting replies. His job prompt names the work, the tool, the categories, the output, and the boundary. This original example uses the same structure:
Review unread messages from the last 24 hours in the approved inbox. Sort them into urgent, informational, and no action needed. Draft replies for urgent messages using the supplied reply examples. Flag missing facts instead of inventing them. Leave every reply for my review; do not send or change messages.
Write the pass condition beside the prompt: which messages count as urgent, which facts every reply needs, and what must be referred to you. Use approved and rejected examples where judgment matters. Brief the model like a brilliant new hire covers choosing that context.
Record how long the task takes you now. During the trial, record the time you spend setting up, running, reviewing, and correcting the agent’s work. Compare the totals to judge whether the automation saves work. Automate the step that slows you down covers choosing a worthwhile first task.
02Limit the tools it can use
A job description tells the agent what you want. Permissions control which actions it can take. In Herk’s email incident, an agent picked up an unassigned task and sent an email intended for human review. The account illustrates two things to check: which tasks the agent can act on and whether it has permission to send.
When connecting the inbox, review the access requested on the provider’s permission screen. Keep only the access needed to read and draft. Check other routes too: blocking one send tool leaves a gap if the agent can use a command-line tool with the same account credentials. Assume your agent uses every tool it has covers the access audit.
In Claude Code, permission rules are enforced by the application. Instructions in a prompt or skill do not change those permissions. Service-side restrictions also matter when the agent can reach an account through several tools.
Herk also recommends using code for fixed rules. A date comparison or a required-field check can run as ordinary code. Reserve model judgment for steps such as interpreting a request. Four shapes an AI workflow can take helps choose how much of the process needs an agent.
Match review to the consequence of an error:
| Action | Boundary to consider |
|---|---|
| Create a private draft | Let the agent finish, then review before use. |
| Send to customers | Require approval before the send, including the recipient list. |
| Change shared records | Limit the records it can edit and preserve a way to restore them. |
| Publish or deploy | Require the relevant checks and a separate approval for release. |
Use the table as a starting point. Set limits according to who an error could affect and whether you can undo the action.
03Check before handing work over
Require the agent to check its output before presenting it. A website needs working controls and readable pages. A report needs supported claims. Ask for the check results as well as the artifact, so you can see what passed and what remains untested. Give your coding agent a finish line covers defining completion.
Use a fresh reviewer when the check calls for judgment. Anthropic’s best practices describe a reviewer that has not seen the conversation that produced the code. Give that reviewer the task, criteria, and output, without the author’s explanation of why the output is good.
For subjective work, first find out whether the reviewer agrees with you. Marchese describes Herk’s support-email example: a person marks 50 replies pass or fail, then the judging model is adjusted against those ratings. Recording yourself explaining each decision can help turn your preferences into usable criteria.
Our recommendation is to keep some examples out of that adjustment process. Compare the judge’s ratings with yours on those unseen examples, and inspect the bad replies it passes. Agreement on the examples used for tuning can overstate how well it handles new work. Score your agent on known good answers covers repeatable evaluation.
Anthropic’s evaluation guide distinguishes checks run by code, models, and people. It recommends checking model graders against human judgments and inspecting the actual result of an action. A plausible completion message does not prove that the intended change happened.
Herk’s review councils give separate agents different audience perspectives. Use those perspectives to find questions you missed. Simulated viewers cannot establish how real viewers will respond, and agreement among models does not replace checking the work.
04Check before expanding the job
Swadia adds calendar access after reviewing the email agent’s work, then schedules the combined task after checking both outputs. Keep each expansion small enough to identify what caused a new failure.
“we don't get promoted easily, neither should your agent”
Sandeep Swadia, “4 AI Agents To Automate 99% Of Your Life” · 6:36Trust applies to the task you checked. An agent that sorts email accurately has not yet demonstrated that it can cancel meetings. Before adding a responsibility, write its pass conditions and test representative cases, including missing inputs and tool failures.
For the inbox agent, review every draft through a trial covering routine requests, missing facts, and tool failures. Before sampling routine drafts instead, confirm that the agent meets the written pass conditions on those cases and reports failures correctly. Keep approval before sending, and resume full review when you find an unacceptable draft or change the model, instructions, tools, or source data. Keep Claude automations working covers testing the actual scheduled run, including its access and failure reports.
05Keep an owner and a run record
Give each automation a named human owner who reads its results and decides when it needs repair. Herk’s feedback advice is to save recurring corrections in a skill. Turn repeated fixes into rules explains that process; Claude Code’s skills documentation describes reusable instructions and supporting files.
Save the failed example and the expected behavior with the correction. Run it again after the change, along with earlier cases that should still pass. A new instruction that fixes one reply can make another worse.
Keep a small record for each automation:
| Field | What to record |
|---|---|
| Owner | Person responsible for checking results and approving changes |
| Job and access | Task, permitted inputs, allowed changes, and approval points |
| Latest reviewed run | Date, output location, and what the person checked |
| Run history | Inputs, tool actions, results, and failed checks needed to investigate an error |
| Failure notification | Where a failed or incomplete run reaches the owner |
Record the last human review separately from the last successful run. They answer different questions. If nobody reads an automation’s output or responds to its failures, reconsider whether it should keep running.
06Start proactive work narrowly
Choose a recurring problem with a clear sign that something needs attention. Herk’s examples include new meeting notes missing from a wiki, conflicting information, and tasks nearing a deadline. He reports better results with these small maintenance jobs than with broad strategic “co-founder” agents. That is his experience, not proof that every strategic use will fail.
Separate finding a problem from changing something. A wiki checker can report a conflict with links to both passages. Rewriting the wiki requires another decision about authority and evidence. Make your agents proactive covers triggers; Keep a growing knowledge base accurate covers conflicting sources.
“I still think there needs to be a human in charge of every automation.”
Nate Herk, “Nate Herk: The Most Valuable AI Skill to Learn In 2027” · 17:31Herk predicts more interaction between team members’ agents. Treat that as a forecast. For agents you connect today, name the human responsible at each handoff and state what the receiving agent may do with a request. A suggestion from another agent should not grant new access or permission to act.
Use this original job card to make those decisions explicit:
Interview me about one recurring task I already do. Ask for missing facts before drafting the job card.
Record the owner, trigger, approved inputs, required output, pass conditions, allowed actions, actions needing approval, and what to do when information or access is missing.
Propose the smallest first run using private drafts or sample data. Separate written instructions from permissions someone must configure. List the checks to run and the evidence the owner should inspect before expanding the job.
Do not connect accounts, change permissions, enable a schedule, or act on external systems while preparing this card.