First pass at delegating errands: weekend plans, cinema tickets, a Blu-ray for the baby, and a mouse price search still running
MiguelX (Twitter) · 2026-09-11
Read originalMuse is a cool way to get started on agentive task delegating - 1st pass: - it made plans for the weekend - bought tickets for Avengers Doomsday - got a copy of Moana on blu-ray for my baby - currently searching for a good deal on a back up gaming mouse for my PC 👏👏👏
Business model
The author calls this his first pass at delegating tasks to an agent. He lists four things: plans for the weekend, tickets for Avengers Doomsday, a copy of Moana on Blu-ray for his baby, and a fourth still in progress - a search for a good deal on a backup gaming mouse for his PC. The post gives no amount, no cinema, no showtime, no retailer, and no description of what the weekend plan looked like. It is a beginner's log; its value is in the way the tasks are bundled, not in any result figure.
The right first task is not the hardest one, it is the one you would not mind getting wrong. This particular bundle is smarter than it looks: suggesting plans costs nothing, buying something you have already named has a fixed target, and only the price search needs a standard defined - and that is the one still unfinished. Scale authority in that order: let it propose, then let it buy what you specified, and only then let it decide what counts as cheap.
Workflow
- Start with the task that costs nothing to get wrong: have it draft the weekend plan first, so you learn what its output looks like and how much you have to edit. No account, no permission, no money - the cheapest possible way to find the edge of your trust.
- Give the plan real constraints: who is coming, the time window, how far you will travel, and a budget. Without these you get a list that would suit anyone and help no one. The more specific the constraints, the more of the output you can use unchanged.
- Then buy only what you have already specified: a disc with an exact title, a ticket for an exact show. The target does not move, so only the execution is delegated. Set a ceiling and the acceptable spec first - edition, language, whether extras matter - and let it find and fill in.
- You click pay. On a first pass especially, submit the order yourself: it checks the amount and the spec, and it also tells you exactly how far down the chain this thing is actually reliable.
- Define what counts as a good deal before any comparison shopping: for an open-ended purchase like a mouse, state the ceiling, the required spec (connection, weight, number of buttons), the brands you will accept, whether refurbished or used is allowed, and how long to wait for a better price before buying at the current one. The item still running in the original post is exactly this kind, and an undefined standard is the likely reason.
- Close with two columns: done, with its artefact, and still running, with what it is blocked on, what is missing, and where to resume. Anything still in progress must never be merged into the completed list.
Adaptation plan
The post gives no amount, no cinema, no showtime, no retailer, no edition and no outcome for the mouse search, and does not describe what the weekend plan contained. Whether the tickets and the disc were actually paid for and delivered is not stated. AgentFlow bought nothing and ran no price comparison.
Prompt
AgentFlow draft · Untested.
You are my delegation assistant. This is our first pass, and the rule is: do the harmless thing first, spend money second, and leave the task that needs judgement last. INPUT - WHAT I WILL PROVIDE 1) A task list, each item tagged: suggest, buy-a-specified-item, or compare-prices. 2) For planning tasks, the real constraints: who is involved, time window, distance, budget. 3) For buying, the exact item name and the acceptable specification, plus a spending ceiling. 4) For comparison, the ceiling, the required specification, acceptable brands, whether refurbished or used is allowed, and the longest I will wait. EXECUTION STEPS 1. Handle only the "suggest" items first. Produce an option set against my constraints and explain the trade-off in each. Do not start any purchase. 2. After I confirm the plan, handle "buy-a-specified-item": find exactly what I named at the spec I gave. Report anything inside the ceiling; stop and explain anything above it. 3. I do the paying and I confirm the order. You find it and fill it in; you do not click pay. 4. Handle "compare-prices" last. First restate my definition of a good deal and wait for me to confirm it. 5. Rank comparison results by total delivered price, listing each candidate and how it differs. If nothing meets the standard, say plainly that nothing met it; do not lower the bar to produce an answer. 6. Finish with two columns: done, with artefacts, and in progress, with the blocker and what is still missing. OUTPUT FORMAT - Plan: time slot / activity / trade-off explained. - Purchases: item / specification / price / order number or artefact / whether I paid. - Comparison: candidate / delivered price / gap against the standard / recommendation (buy, wait, walk away). - Two closing columns: done and in progress. HUMAN CONFIRMATION I make every payment and every order confirmation. The plan must be confirmed by me before any purchase begins. My definition of a good deal must be accepted by me before any comparison starts. FAILURE AND MISSING DATA If a constraint is missing, stop and ask rather than delivering a generic plan. If a price exceeds the ceiling, stop and explain. If nothing meets the standard, say so instead of relaxing the standard to produce a result. Retry a given task at most once; if it still fails, list it as in progress with the blocker.
Running costs and pitfalls
The post gives no amount, no cinema, no showtime, no retailer, no edition and no outcome for the mouse search, and does not describe what the weekend plan contained. Whether the tickets and the disc were actually paid for and delivered is not stated. AgentFlow bought nothing and ran no price comparison. No independent cost benchmark is available. Check subscriptions, permissions and external actions before running a workflow.