Ilya Arbabi

An agent is fixing a bug and starts touching files outside the requested scope. Before switching tools, I'd check whether the task has a reproducible failure and a clear boundary. Without those, the next agent gets the same ambiguity.

If the task is well defined, I'd give Codex and Claude Code the same starting commit and stop at the first reviewable diff. I'd compare unrelated edits, tests actually run, and the human work needed before merge. A passing test alone misses that cost.

If you've switched mid-task, what triggered it? Did the second tool make the patch easier to review, or did you mostly improve the brief on the second try?

WMM: my AI agent does the marketing, I still sign off on what ships

I build apps and handle their marketing myself. WMM is a local marketing workspace my own AI agent operates: it researches, drafts and schedules the work inside permissions I set per operation. The actions I choose to gate wait for my explicit approval of the exact proposal before anything runs, and when a reviewed action executes it leaves a saved receipt of what actually ran.

Honest state: public repo, one documented trial on my own app's Facebook Page - I approved the content, a scheduled run published it, and the readback from Meta matched what I approved. One post. Nobody is retiring off this yet.

What I want from this crowd: if your agent queued a week of marketing work for your review, what do you need to see per item before you'd approve - content, reason, destination, cost? And what would you never take out of the queue?

So, been running a AI startup for more than a year now and the one thing i always thought or was puzzled was when naturally customers think AI should be cheaper than human hours even though AI's throughtout or the benefits is 100x.
So we have been doing it the 'cost+' pricing way, but to be very frank this is working and is helping us be in control of the margins, but also exploring outcome based pricing in certain use cases. Will have this thread open to have discussions around this topic, also if i find newer ways or use cases where outcome based pricing works, happy to explore.