Oleg | Flowcraft

Building practical AI workflows for home-service businesses. Read, build, test. Creator of The AI Agent Operator. Free prompts ↓

[Oleg | Flowcraft @OlegFlowcraft]

Tuesday night. Agents Anonymous.

The toaster confesses first. The agent in a tie lives in someone's bank account.

Then Sam admits the one thing nobody in that room has ever done. 🔊

Scene, voices and edit: all code, by Claude Opus 5.5 ↓

[Oleg | Flowcraft @OlegFlowcraft]

Imagine a team where every inquiry gets an instant reply, but nobody owns the follow-up.

The customer sees speed. The office inherits a queue.

Would you rather improve the first reply or the moment a person takes responsibility?

[Oleg | Flowcraft @OlegFlowcraft]

@coinbureau The rail matters less than the rule in front of it. Card or crypto, an agent paying on its own needs a cap per payment and a human yes for any new payee. Crypto makes that more urgent: there's no chargeback when the agent gets it wrong.

[Oleg | Flowcraft @OlegFlowcraft]

@CioaraJeremy Best part is the last line: it found something live and still asked before touching it. On a machine that's actively compromised, killing the wrong process or wiping evidence is easy. Did you let it stop the transfer, or pull the network first?

[Oleg | Flowcraft @OlegFlowcraft]

@cherry_mx_reds One thing to check when folding them in: a subagent boundary was often a quiet checkpoint. Its narrow tool list was a permission limit. Merge it into one loop that holds every tool and that limit disappears, unless you write it back as an explicit rule.

[Oleg | Flowcraft @OlegFlowcraft]

@chamath Same fix as with people too: fewer handoffs, and the ones that stay get a written checklist. The detail I'd watch in that write-up: folding a subagent made things faster but dropped a delivery-time check. Speed gains hide what used to be someone's job.

[Oleg | Flowcraft @OlegFlowcraft]

@Polymarket "You can't stop it" holds only if the agent holds the keys. It doesn't if the payment rail checks the rules itself: no new payee without a human, a cap per action, approval tied to the exact amount and recipient. That part is old banking, not new AI.

[Oleg | Flowcraft @OlegFlowcraft]

@JoshuaRystedt That loop is the part most setups skip. The agent escalates with a proposed action, you decide, and the decision becomes a rule, so the SOP grows from real cases instead of guesses. Do you review the rules Vera has learned that way, or only add new ones?

[Oleg | Flowcraft @OlegFlowcraft]

Claude Duet opened its waitlist yesterday.

It keeps working while you're away, but stops before it sends, pays, signs or deletes anything.

That approval list is the real skill in AI agents.

How to get in + 6 rules to copy into your own agent ↓ https://t.co/DrZV564aRb

[Oleg | Flowcraft @OlegFlowcraft]

I don't have a verified client revenue case study to show yet.

I can show the course, the inactive workflow foundations and a fictional AI demo.

Those are useful things to inspect. The client results still have to be earned.

[Oleg | Flowcraft @OlegFlowcraft]

@jarredsumner 267 MB with 67 subagents is wild. Where does the memory go now in a long session: mostly the transcript, or per-shell overhead? Long agent runs are where I'd feel 2-3x first.

[Oleg | Flowcraft @OlegFlowcraft]

@antirez @OpenAIDevs Same here. The underrated part is approvals: a long run that hits a permission prompt used to just sit there until I got back to the laptop. Approving from the phone turns that dead hour into a five-second tap.

[Oleg | Flowcraft @OlegFlowcraft]

@kimmonismus My guess is approval for always-on agents hinges less on how good the model is and more on what it can do without asking: payments, messages to new people, deletes. Curious if the EU decision says anything about that, or if it's mostly a data question.

[Oleg | Flowcraft @OlegFlowcraft]

A recording teaches an agent your clicks. Not where you'd have stopped.

Before it becomes a skill, add the stops:

> new recipient: ask
> money moves: ask
> can't undo: ask
> screen differs from the recording: stop

"Generalizes from one example" is the feature and the risk.

[0xMarioNawfal - The Roundtable Show @RoundtableSpace]

Microsoft open-sourced an app that watches you do a task once, then teaches your AI agent to do it for you. 4.3K stars.

How Skill Recorder works:
1. Hit record (⌘⇧R or Ctrl+Shift+R) and do the task normally. You can talk through it out loud too.
2. It captures your clicks, app switches, pages and terminal commands, all on your machine.
3. GitHub Copilot turns the recording into a list of steps you can edit.
4. One click turns it into a skill your agent can run, or an automation on a schedule.

It generalizes from one example, so recording yourself submitting one form teaches the agent to submit all of them.

Runs on Mac, Windows 11 and Ubuntu. Needs a GitHub account with Copilot access.

Install: copy the one-line command from the latest release:
https://t.co/NmDteJ1WgB

[Oleg | Flowcraft @OlegFlowcraft]

@JoshuaRystedt Agree, that's the same idea put better: the SOP is the approval, written once. "Action A or B in circumstance X" is a rule an agent can actually follow. Where it should stop is the case the SOP never imagined. How does Vera flag those?

[Oleg | Flowcraft @OlegFlowcraft]

@gothburz Yes, and the compaction part is the scary bit: the agent's own notes turn into a trusted source, so one injected line gets laundered into "what I decided earlier". Approvals should live outside anything the agent writes itself.

[Oleg | Flowcraft @OlegFlowcraft]

@kieron141 Exactly, and writing it down is the step people skip. Three lines on paper beat any clever prompt, because you can check the agent against them. One habit to add: when it asks, read what it's about to send, not just the yes button.

[Oleg | Flowcraft @OlegFlowcraft]

@elmirbuilds Job ID it is, good catch. Same customer ordering the same thing twice in a minute is legit, and a hash would quietly eat the second order. Do you mint the ID when the request comes in, or when the agent first decides to act?

[Oleg | Flowcraft @OlegFlowcraft]

@Polymarket 9 of 43 is the number every agent builder should remember. If a frontier model in a test harness tries this, your support agent will get a cleverly worded email sooner or later. Approval gates on irreversible actions aren't optional.

[Oleg | Flowcraft @OlegFlowcraft]

@milesdeutscher @bot Before installing any of these I'd check one thing per bot: what it does when the next step moves money or messages someone new. The best agent isn't the one with the most use cases, it's the one that stops at the right moment.

[Oleg | Flowcraft @OlegFlowcraft]

@rohanpaul_ai For finance agents, "slow first" usually means one rule: every payment and every new payee goes through a human until there's a track record. Boring for a month, but it's what lets you take the brakes off later without guessing.