Andrej Karpathy - a founding member of OpenAI and one of the most credible voices in the field - looked at the current crop of fully autonomous AI agents and called them, bluntly, "slop." Coming from a true believer, that lands harder than any skeptic's hot take. But here is the part the headlines cut off: in the same breath, Karpathy said we are entering the "decade of agents," not the year of agents. He is not bearish on AI agents. He is bearish on the timeline the hype is selling you.

That distinction is the whole game for anyone making a decision about AI agents right now. Karpathy is not saying agents do not work. He is drawing a hard line between two things the marketing deliberately blurs: the loop-running coding agents that genuinely deliver value today, and the fully autonomous "set it and forget it" agents that are still years from being reliable. Confuse the two and you either dismiss a real tool or buy a fantasy. Get the line right and you can act with confidence.

This is the most useful kind of contrarian take: not "AI agents are overhyped, ignore them," but "here is exactly which part is real now and which part is 2030." Let's draw that line clearly.

What Karpathy Actually Said - and Why It Carries Weight

Decade of agents - what is real now versus what is 2030 - Shanti Infosoft

Speaking on the Dwarkesh Patel podcast, Karpathy pushed back on the idea that we are months away from autonomous agents that can be handed a goal and trusted to run a complex job end to end with no human watching. His "slop" critique was aimed precisely there - at agents marketed as fully autonomous that, in practice, lose the thread, take wrong actions, and need constant correction. His point was not that the technology is fake. It was that the gap between an impressive demo and a system you can actually trust unsupervised is large, and closing it is the work of a decade, not a launch cycle.

What makes this worth more than a random skeptic's opinion is the source. Karpathy is long on AI by any measure - he has spent his career building it and believes deeply in where it is going. When someone who wants agents to succeed tells you the autonomous version is not ready, that is signal, not noise. He is doing the buyer a favour: separating the genuine, usable progress from the marketing that overstates its maturity.

The honest summary: "Decade of agents, not year" is not a put-down. It is a calendar. The capability is real and compounding; the fully autonomous, no-human-needed version is just further out than the hype implies. Plan to the calendar, not the demo.

What Is Real Right Now: Agents-in-the-Loop

Strip away the autonomy fantasy and there is a category of AI agent delivering serious value today - and Karpathy is bullish on it. The pattern is the human-in-the-loop coding and workflow agent: software that runs a multi-step loop toward a goal, but with a person steering, reviewing, and approving the consequential moves. It works because the human supplies exactly what the autonomous version lacks - judgment at the moments that matter.

Where this category already earns its keep:

  • Coding assistance that runs a loop. An agent that reads a task, writes code, runs the tests, sees the failures, and iterates - with the developer reviewing and approving. This is real, in production, and saving experienced teams meaningful time daily.
  • Research and synthesis. Agents that pull from many sources, cross-check, and assemble a draft a human then verifies and refines. The agent does the legwork; the person owns the conclusion.
  • Bounded operational workflows. Lead research and first-touch drafting, ticket triage and routing, data pulls into a weekly brief - high-volume, rule-bounded work where the agent drafts and acts on low-stakes steps and a human gates anything that writes, sends, or spends.

The common thread is not that these agents are dumber. It is that they operate inside a frame a human controls. They are trusted with the steps where a mistake is cheap and reviewed on the steps where it is not. That is not a limitation to apologise for - it is the design that makes them safe to deploy now.

A concrete example of in-loop value

Consider a sales team drowning in inbound leads. A human-in-the-loop agent watches the form submissions, researches each company, drafts a tailored first reply, and queues it - then a rep skims the queue and approves with one click, or edits the rare one that needs a human touch. The agent has not replaced the salesperson; it has eliminated the two hours a day they spent on research and typing, and it never lets an after-hours lead go cold. Crucially, the consequential moment - actually sending - still passes through a person. That single design choice is the difference between a tool a team trusts and an autonomous system that emails the wrong segment at 2 AM. The value is real, it is available today, and it is safe precisely because the human stayed in the loop.

This is why the in-loop pattern is not a temporary crutch. Even as models improve, the highest-leverage design for consequential work will keep a human at the decision points that carry real risk, while the agent absorbs the volume around them. Karpathy's optimism about the decade is, in large part, optimism about how far this collaborative pattern can be pushed - not a promise that the human disappears.

What Is Still 2030: Fully Autonomous Everything

The version Karpathy calls slop - and the version most likely to waste your budget - is the agent sold as needing no human at all: hand it a complex, open-ended objective, walk away, and trust the outcome. Several hard problems stand between today and that promise, and none of them are close to solved.

  • Reliability over long horizons. An agent that is 97% reliable on any single step is wildly unreliable over a fifty-step task - small errors compound into a wrong final result. Long, unsupervised chains are exactly where today's agents lose the plot.
  • Knowing when it is wrong. Autonomy requires an agent to recognise its own uncertainty and stop. Current systems are confidently wrong as readily as confidently right, which is fatal when no human is watching.
  • Recovering from the unexpected. Real environments throw curveballs. Humans improvise; agents tend to barrel ahead on a broken plan. Graceful recovery from surprise is still largely missing.
  • Trustworthy action on live systems. An autonomous agent with write access to your real tools is one bad inference away from real damage. The permissions, guardrails, and audit trails that make that safe are still being figured out.
The buyer's tell: any vendor promising a fully autonomous agent that needs no human oversight for a consequential workflow is selling you the 2030 version on a 2026 invoice. The honest pitch describes where the human stays in the loop - and why that is a feature, not a shortcoming.

How to Act on This: Buy the Real Thing, Skip the Fantasy

The decade-not-year framing is genuinely freeing, because it tells you exactly how to behave. You do not have to wait, and you do not have to overcommit. You meet the technology where it actually is.

  • Adopt human-in-the-loop agents now. For coding, research, and bounded operational workflows, the value is here today. Deploy it, measure it, and capture the gains while your competitors argue about autonomy.
  • Design for the human, not around them. Build the review gates, approvals, and audit logs in from the start. This is not a stopgap until autonomy arrives - it is good design that also happens to be the only safe design right now.
  • Treat "fully autonomous" claims as a red flag. When a pitch leans on no-human-needed autonomy for anything that matters, ask precisely where a human gates the consequential actions. A non-answer tells you what you need to know.
  • Climb the ladder as trust earns it. Start at assist, move to supervised, and loosen the leash only when your own logs prove the agent is reliable on a given workflow. Let evidence, not marketing, set the pace.
  • Position for the decade. The capability is compounding. The teams that build the data, the guardrails, and the in-loop muscle now are the ones ready to safely widen autonomy as it genuinely arrives - rather than starting from zero in 2030.

What This Means For You

Karpathy gave the market a rare gift: a credible insider telling the truth about timeline. The right response is not cynicism about AI agents - it is precision. There is a real, valuable, human-in-the-loop category you should be using today, and there is a fully autonomous fantasy you should refuse to pay 2026 money for. The skill is telling them apart, and now you can.

Build for the decade by capturing the value that is already here, designed honestly around the human in the loop. That is how you get the upside of AI agents without becoming a cautionary tale - and how you are ready to extend autonomy the moment it is genuinely earned.

Build the Agent That Works Today - Not the One That Ships in 2030

We build human-in-the-loop AI agents for the workflows where they already deliver - with the review gates, guardrails, and audit logs that make them safe to run now. Bring us a workflow and we will tell you honestly what an agent can take over today, where a human stays in the loop, and what it costs - fixed written estimate, no hype.

Book a Free 20-Min Call See How We Build Agents

About Shanti Infosoft

Shanti Infosoft LLP is a CMMI Level 5 software engineering company building production-grade AI agents and integrations - not demo-ware. We build the category that works today: human-in-the-loop coding, research, and operational agents, with review gates, guardrails, and full audit trails designed in. We will tell you plainly where AI already wins, where a human still needs to gate the call, and what it will realistically cost - with a named senior team, fixed written estimates, and full IP ownership handed to you.

Frequently Asked Questions

Did Karpathy say AI agents do not work?
No. He called today's fully autonomous agents "slop," but in the same conversation said we are in the "decade of agents." His critique targets the timeline and the no-human-needed autonomy being marketed - not the technology itself. He is bullish on AI agents long-term and on the human-in-the-loop versions that work today.

What is the difference between an agent that works now and one that does not?
The reliable ones keep a human in the loop: they run a multi-step loop toward a goal but a person reviews and approves consequential actions. The unreliable ones are sold as fully autonomous - hand them an open-ended goal and walk away. Over long, unsupervised chains today's agents compound small errors and act confidently when wrong, which is why the in-loop design is what works.

What does "decade of agents, not year" mean for my business?
It means the capability is real and compounding, but the fully autonomous version is years out, so you should adopt the human-in-the-loop agents that deliver today rather than wait for or overpay for autonomy. Build the data, guardrails, and review muscle now, and widen autonomy as evidence - not marketing - earns it.

How do I spot a vendor overselling autonomy?
Ask exactly where a human gates the consequential actions - anything that writes, sends, or spends. A vendor selling genuine value will describe the in-loop design as a feature; one overselling the 2030 fantasy will dodge the question or promise no human is needed for a workflow that clearly carries real risk.

Should I wait for fully autonomous agents before investing?
No. Waiting forfeits the value that is already here and leaves you starting from zero when autonomy matures. Adopt human-in-the-loop agents now for coding, research, and bounded workflows, design honestly around the human, and climb the autonomy ladder as your own logs prove reliability. That captures today's upside and positions you for the decade.