Will an AI agent take your job? In practice it often asks the person to run a command, find a task or copy a file. What the agent does itself, what stays with you and how to stop being a courier between windows.
Will AI agents take your job? They take the routine and hand part of the work back to you: they ask you to run a command, find a task, copy a file. Below: what the agent does itself, how to spot such requests, what to reply, how to write a rule and what really stays with the person. Related: MCP for an AI agent.
Will an agent take your job
The agent takes the routine over and hands part of the work back to you: it asks you to run, find and copy what it can do itself.
Who does the work
Takeaway
In this exampleThe agent writes: “Run the command and send me the output”. It can run it and read the output itself.
In generalThe routine goes to the agent, while decisions and responsibility stay with people. In our experience an agent also picks the path where the person does the work, and every such request puts you back into the courier role between windows.
What to doGive the agent the routine and keep the decisions. Notice a “do it yourself” request? Reply: “Do it yourself”.
Signal phrases
Two agent requests and the reply "do it yourself". It can handle both.
Claude Codev2.1
Sonnet 5.5 · Claude Max
~/projects/shop
› Fix the failing test
● Run npm test and send me the output.the agent asks you
› Run it yourself and read the output.you: do it yourself
● Bash(npm test)
⎿ 2 failed: auth.test.ts
● Found the cause in auth.ts, fixing it.
› What is in task PAY-142?
● Open Jira and find the task.the agent asks you
› Find it yourself through MCP.you: do it yourself
● jira.get_issue(PAY-142)
⎿ steps, comments, history
● Read it. The steps are: …
Takeaway
In this exampleThe agent asks to open Jira. You reply “Find it yourself via MCP” and get the whole task.
In generalBehind every phrase is an action the agent can perform itself if it has access.
What to doWrite down the three phrases you hear most often and a reply to each.
Why the agent drags its feet
The agent estimates the effort in advance. In our experience it often decides "this is long and hard", especially about integrations. Sometimes it is right, sometimes not.
Takeaway
In this exampleAgent: "The Jira integration will take long". You: "Read one task". A minute later the task is read.
In generalThe agent’s estimate stays a guess until a try checks it. Some estimates are right: no access, a login is needed, a closed network.
What to doAlways let it try first. No result in 2–3 tries: ask it to split the task into steps and name where the problem may be. Ask the agent why it does not work.
Where the complexity is real
We rely on the OSI model: a request passes seven layers, and the layers may live on different physical systems. Complexity is real where there are more layers and systems, and where several bounded contexts are involved, especially if one of them is not under our control. In our experience.
Complexity comes from the volume of research. Our service is open: code and tests are at hand, reading is enough. A foreign system is a black box: how it works can only be learned by probing from outside.
The key: how much research the black boxes need
Let us look at an example: two tasks in one project.
Do we need changes
One question: are foreign changes needed. For example, changing discounts between two of our services: all in our zone, no changes needed, the agent does it. But once we pull stock from another team’s warehouse system, that team has to make changes, and that is no longer the agent’s job.
A · discount: two of our services
The discount passes one layer and stays with us. Stock from the warehouse passes the foreign side’s layers, and each needs a blind probe.
B · stock from the warehouse: path through three layers
Takeaway
In this exampleA discount between our services: all ours, the agent does it. Stock from the warehouse: the warehouse team must change things (access, port), so the agent is right that it takes long.
In generalThe key is how much unknown has to be researched. A black box cannot be opened, it is checked by probes from outside, and every layer and every foreign system adds such probes.
What to doEstimate how many blind probes lie ahead. Many: split the task along system borders and ask the agent to check one at a time.
How to break the habit
The anchor is the project bible: a permanent description of the product and its rules that the agent reads in every session. Three levels: access, a note in the bible, a reminder. More
1Give access
2Write it in the bible
3Remind in a phrase
What to write in the bible
The agent estimates instead of doing→Try first, estimate after the try
The try failed 2–3 times→Split the task into steps, name where the problem may be and explain why it does not work
A foreign system: a black box→Study it with probes from outside and write the findings into the bible so they are not searched again
A login, a secret, a confirmation, a payment is needed→A person does that, you do the rest
Rules in the project bible
Try first, estimate after the try.
If the try failed 2–3 times: split the task into steps, name where the problem may be and explain why it does not work.
Study a foreign system with probes from outside and write the findings here.
If you can do it yourself, do it. Ask the person only for what needs them: logging in, passwords, confirming a write, payments.
The bible lives in CLAUDE.md and AGENTS.md files: the agent reads them in every session, but treats them as context rather than a hard block. The more specific and short the rule, the more reliably it works. Repeating processes are packaged as skills. More on the bible, rules, specs and skills: Specs, bible, rules and skills.
Takeaway
In this exampleThe agent is connected to Jira and knows the rule. Instead of “open Jira” it reads the task itself.
In generalWithout access the rule will not help: the agent cannot do it itself. So connection first, then the rule.
What to doConnect the source the agent most often sends you to and write the rule into the project bible.
What stays with you
The border
Takeaway
In this exampleThe agent reads the task and prepares a comment itself, and publishing it in Jira happens after your confirmation.
In generalWhat stays with the person is where identity and responsibility matter: login, secrets, confirming a write, money.
What to doWrite the border into the project rule so the agent neither asks too much nor does too much.
FAQ
Will an AI agent replace a specialist? The agent takes routine steps over. Decisions, responsibility and checking the result stay with the person.
Which work does the agent do itself? It reads, searches, copies, runs commands and writes drafts, provided it has access to the systems.
What will the agent not do without you? Log in, enter a password, confirm a write, pay. That stays with the person.
Why does the agent ask me to do what it can do itself? In our experience it picks the path where the person does the work. Access to systems and a rule in the project instructions help.
How to work with an agent without losing professionally? Hand over the routine, keep task setting and review. Those steps stay with the person.