AI Guides · Natthan.AI Lab
AI agents vs AI assistants: who carries the work?
Use the Natthan.AI Lab's night runner and held releases to decide whether you need an assistant or an agent. Write down the work you still carry.
Last updated
What you'll be able to do
- Decide whether your tool carries work between prompts, using the night runner's refusal in Note 015 as the test.
- Write what stays with you, or a job with a stop someone else controls, using the lab's published roles as your model.
Who carries the work?
Start with the terms page's distinction. An assistant helps you finish a task you hold. An agent carries a defined area under a charter, a written job with permissions and stop conditions, and returns a receipt, evidence you can check against the actual result.
Keep the director role, the human responsibility for direction and acceptance, in either arrangement. Apply that split even when an agent runs without you watching, as the content runner did in Note 015.
Give delegated work a gate, a check that can stop it, held by someone other than the builder. Use the release held after independent review in Note 012 to test whether that stop has any force.
Run the test on your own tool
1. Ask the three questions
Pick the tool you used most this week. Open its last completed job and answer from the record, using these Lab Notes as checks.
Who holds the outcome when nobody is watching?
Look for a heartbeat, a scheduled run that checks whether work is due. Compare what happens between your requests with Note 015: the night runner declined a sixth same-day post and checked what was already live. Mark the work as yours if it waits for you to choose and start the next task.
What evidence comes back unprompted?
Open the receipt that arrived with the job. Try to catch the blank page in Note 004 using only what your tool returned. Count a completion message without checkable evidence as a reporting gap you still have to close.
What can stop it?
Name who can hold the work when the builder says ready. Compare that authority with the built-right and will-it-earn vetoes in Note 011, where neither side can override the other. Keep yourself at the stop when you use an assistant; give an agent a separate gate with the consequence shown in Note 012's held release.
You know it worked when
You can point to who advanced the last job between requests and open the evidence it returned. You can also name who could have stopped it, as the reviewer did in Note 012. Mark missing evidence or a missing stop as unfinished delegation, even if the tool runs on a schedule.
2. If it's an assistant, write what you keep
Write “What stays with the owner” above the part you still carry. Use the lab's director role to name the decision you keep and the result you will inspect. For the content job below, keep the publishing decision with you and check the draft before it goes live.
You know it worked when
You have written who will open the published page before calling the job complete. Use Note 004's blank-page report to check that your plan would catch missing work after a convincing draft or status message.
3. If it's an agent, write its charter line and gate
Copy the scope of Lenny's published role: responsibility for a named lending marketplace. Write your own line in that shape: “This agent owns [area] and may [permitted action] until [stop condition].” Keep it within the lab's charter definition.
Put the gate holder beside that line. Use Codex's independent review role as the model for a release check outside the builder's control. Require a failed review to hold the release, as it did in Note 012.
You know it worked when
You can point to an enforced stop as well as the written job. Mark any unwired rule as pending, following Note 016's distinction between written laws and working gates. Keep the affected action under your control until that stop works.
4. Pick the smaller arrangement that does the job
Apply the Fit / Stretch / Trap screen. Check whether the job has familiar patterns and whether errors are cheap or checked. Ask whether you are using AI to think or handing it your judgment.
Keep a trap, a job missing grounding or a check on consequential errors, with you; use an assistant only to help examine it. Treat a stretch as work needing deliberate verification. Consider an agent for a fit that recurs and has a result you can inspect, like the bounded content work in Note 015.
You know it worked when
You can explain why this job needs another run without a fresh request, or why you will keep it as a conversation. Allow a justified refusal to count as useful work, as it did for Note 015's repetitive post. Keep the assistant if you cannot yet define the agent's stop.
Worked example
Compare the same content job both ways. Use the account published on , the night runner's refusal to publish, for the agent side.
As an assistant
Ask for a post and get a draft. In this version of the job, I still decide whether another post belongs on the site. I carry the draft through publication and check the result. The next useful action waits for me.
As an agent, from Note 015
My night runner declined a sixth same-day templated post on Aug 3. Four had already shipped that day, across 16 trade verticals. It ran the health checks, read the demand data and the one post already covering the strongest topic, and drafted nothing.
Put yourself in my morning review: read the refusal as a receipt for the publishing decision. Check the changes to the live posts separately before accepting that work as complete. Use the specific reasons in Note 015 to distinguish a deliberate refusal from an empty status report.
Mistakes to avoid
- Calling a report a receipt. Open the result before accepting the claim. Check for the failure in Note 004: a blank page called complete and a safety mechanism called armed before it was wired up.
- Letting the builder clear its own stop. Keep the built-right and will-it-earn vetoes separate, as in Note 011. Reproduce disputed findings too: the independent auditor in Note 012 also reported build failures that clean reruns disproved.
- Paying attention only to new output. Review the reason for doing nothing alongside shipped work. Use the declined repetitive post as your check that the agent may protect the content area by refusing more of the same.
- Treating written rules as enforced gates. Label an unwired rule honestly. Follow Note 016, which records that some laws had working checks while others were still written law.
Ten minutes, today
Open the tool you used most this week and run the three questions on its last job. Write “assistant,” “agent,” or “unfinished delegation” beside the evidence. Use the night runner in Note 015 and the held release in Note 012 to challenge that label.
Finish with what stays with you. Keep the publishing decision if you are using the assistant version of the content job; name the independent gate if you are handing over the recurring work. Leave with that line written before asking for the next draft.
You know it worked when
You have a label backed by the last job's evidence and a named person who accepts the result. Check that final responsibility against the lab's director role.
Questions about agents and assistants
What's the difference between an AI agent and an AI assistant?
Use the content refusal in Note 015 as my dividing line: the runner chose useful work without waiting for another prompt. Compare that with asking for a draft and carrying it to publication yourself. Apply the terms page's distinction to the work your tool actually carries, whatever its product name.
Isn't 'using AI tools' the same as running an AI company?
Look at the split between Lenny and Codex on the Example team page. I give Lenny the lending marketplace to run; Codex has an independent review role. Compare that division of work with a tool collection where you still start every job and carry every handoff. Keep human accountability in either arrangement, as the lab's AI-agent-run company definition requires.
Why does the distinction matter?
Try checking the blank-page completion claim in Note 004 from the status message alone. You still have to open the result to discover that the work is missing. Require that evidence to come back with the job before you count it as delegated. Keep the failed-review stop in Note 012 outside the builder's control.
When is an assistant the right call instead?
Use the lab's Fit / Stretch / Trap screen before handing over a job. Keep judgment with you when the facts or checks are missing; conversation can help you examine that gap. Take Note 012's false audit findings as a reason to reproduce the disputed result before accepting either agent's conclusion.
Read the held release and disputed audit findings in Note 012.