Contrarian take to start: the world is not ready for AI agents, and the fact that OpenAI’s head of product says otherwise is exactly why I don’t believe it. That’s not a shot at Thibault Sottiaux personally. It’s a comment on the job description. Heads of product are paid to say the market is ready. If the market weren’t ready, they’d be saying the market is “early” and “learning fast,” which is the same sentence wearing a different jacket.
Here’s what we actually have from the TechCrunch interview: Sottiaux said the world seems to be ready for AI agents, and he talked about cost and efficiency. The piece framed adoption as moving quickly. That’s the whole verified core, and I’d rather work with a thin set of real facts than a thick set of imagined ones.
What “ready” means when a product lead says it
Readiness is not one thing. It’s at least four, and they mature at wildly different speeds.
- Model readiness — can the thing reason through a multi-step task without falling over? Improving, visibly, month over month.
- Tooling readiness — do the harnesses, permissions, logs, and rollback paths exist? Patchy. Very patchy.
- Organizational readiness — does anyone at the company own the agent when it does something dumb at 2am? Mostly no.
- Economic readiness — does the task cost less to run as an agent than as a human plus a script? Depends on the task, and nobody publishes the losers.
When a product leader says “the world seems to be ready,” they’re usually talking about the first one and betting the rest catches up. From where I sit testing these toolkits every week, the gap between model quality and everything else is the single biggest source of disappointment users report to me. The model is rarely the weak link anymore. The plumbing is.
The cost conversation is the interesting part
Sottiaux discussing cost and efficiency is, to me, the part worth chewing on. Agents are the first AI product category where the unit economics are visible to the person using them. With a chatbot, you pay a flat monthly fee and the provider eats the variance. With an agent, you feel every retry. You feel the loop that ran eleven times because a selector changed. You feel the tool call that returned a 40,000-token blob of HTML the model then had to read.
This changes how you should evaluate any agent toolkit. Speed and demo polish are cheap signals. The signals that matter:
- Does it show you token spend per run, not per month?
- Can you cap a run before it burns your budget?
- Does it cache anything, or re-read the same context on every step?
- When it fails, does it fail loudly and stop, or quietly and continue?
Most toolkits I’ve reviewed handle the first two badly and the last one worst of all. An agent that quietly continues after a failed step is not an agent. It’s a random number generator with API access.
Rapid adoption is not the same as rapid success
The TechCrunch framing leaned on how fast adoption is moving. I believe it. I also think adoption numbers are the least useful metric in this category, because trying an agent costs almost nothing. Sign up, paste a key, watch it do something impressive, tell your team about it. That’s adoption. Whether it’s still running your workflow in ninety days is a different question, and it’s the only question I care about as a reviewer.
My rough read from testing: the tools that survive ninety days are the ones that do one narrow thing with a tight feedback loop. Code changes with a diff you review. Research with citations you can check. Data cleanup where the before and after are both in front of you. The tools that die are the open-ended “give it goals and walk away” pitches, because nobody can tell whether they worked.
So is the world ready or not
Partly. And “partly” is a real answer, not a dodge. The models are good enough that agents are no longer a research demo. The economics are getting there, which is presumably why an OpenAI product lead is talking about cost in an interview rather than capability. That’s a tell, and a positive one. Companies talk about cost when capability stops being the bottleneck.
What’s missing is boring infrastructure. Audit trails. Permission scoping that a non-engineer can configure. Cost ceilings enforced by the platform rather than by hope. Clear ownership when an agent takes an action someone has to explain to a customer. None of that is glamorous, none of it makes a good conference slide, and all of it is what separates a tool people try from a tool people keep.
Sottiaux said the world seems to be ready. That hedge is doing honest work. I’d take the same position with one addition: the world is ready to use agents, and mostly not ready to be responsible for them. Fix the second half and the first half stops being a debate.
🕒 Published: