TheInformation published a report indicating OpenAI’s plans to rollout AI agents where the newest novelty is their ability to demand a human wage. With OpenAI aiming for the agents to generate 20-25% of long-term revenue and OpenAI investor SoftBank having committed to $3 billion on OpenAI’s services this fiscal year, the world will quickly find out: Can AI agents replace human workers? Perhaps we need to look no further than the 298 job postings currently listed on OpenAI’s Careers page to find our answer, but this article will not stop there to peer beyond glossy veneers and investigate a more nuanced story, anchored against observed AI constraints, inevitable human oversight requirements, and security considerations which will shape the adoption of the premium agent tiers.
OpenAI’s agent offerings will target three distinct professional domains:
| Price (per month) | Agent |
| $2,000 | Knowledge Worker |
| $10,000 | Software Developer |
| $20,000 | PhD-level Researcher |
It is not confirmed but reasonable to expect that the agents are developed in a way to allow for each subsequent tier to complete the same work as previous tiers seeing as all agents will leverage GPT 4.5 (an intermediate upgrade to 4o), currently underwhelming ChatGPT Plus users in Research Preview. The key difference may lay in increasing context windows with the resources to reach increasingly nuanced conclusions across widening domains in reasonable time windows.
This development may suggest a new brute force approach from OpenAI, where they acknowledge an impending performance plateau for current generative AI models. In fact, it was likely with the reasoning models that OpenAI first acknowledged this in subtext on the public stage. While the organization shows no signs of stopping research, with a $500 billion joint venture investment for AI infrastructure, they do seem to increasingly throw more money, GPU clusters, and data at a problem which needs novel, publicly shared research to drive it forward. After all, if OpenAI cannot get the newest model to stop hallucinating roughly a third of the time, why not have the model check itself over-and-over until a greater reliability can be proven?
Even at a $20,000 price point, current AI systems have fundamental limits where the model cannot only make mistakes but inevitably will, overlooking important nuances and getting stuck on problems which professionals could navigate with ease. Humanity possesses something current generative models never will: Ingenuity. It’s remarkably human, and at these price tiers the market will no doubt demand the same from these AI until we can learn a collective lesson. Without a human “on call” to review and fix stuck points and ensure alignment of the agent, even premium tiers will bring projects to a halt at precisely the worst moments.
The current approach by OpenAI for agents has been continual human supervision where the agent always reports back to the user directly; however, with the introduction of Operator, the organization seems poised to contribute to the inefficient, overly verbose, and poorly optimized outputs and the dead internet which define the broad criticism of AI at present. Experienced developers may recognize some regular flaws listed below:
This content, which some online have deemed AI slop, requires significant human clean-up and means expensive AI agents still won’t deliver the “plug-and-play” or acceptable baseline productivity which companies will demand at the price point unless a supervisor or team of supervisors are “in-the-loop.”
The most pragmatic application of OpenAI’s premium agents—particularly the $10,000/month developer tier—may be as Pull Request (PR) agents within larger, established development teams with enough content generated to shape the agent’s output. Rather than replacing developers outright, the agents would serve as overnight assistants to handle more routine coding tasks and updates, only proposing initial implementations to prepare PRs for human review and iteration each morning and throughout each day.
As we have no specific details published by OpenAI, the below sub-section is entirely speculative.
The developer agent’s substantial price tag may make economic sense in some contexts. The following scenarios may be those where the agent finds itself transformed from the unrealistic “developer replacement” some fear it to be to a specialized tool designed to handle more predictable, routine aspects of software development. Perhaps this may allow human developers to retain their job, focusing on complex problem-solving, architecture decisions, and code review which make the job enjoyable.
As someone with GitHub Copilot and Cursor experience for personal coding projects intended for the cutting edge, I have never been impressed with either and find myself regularly rolling up my sleeves to “get things done,” even with the expensive o1 Pro. More autonomy to these agents may reduce this need, but I am entirely skeptical that they would eliminate the need entirely.
As companies bring AI agents into their workflows, there’s another critical consideration: Security and access controls. With the context of previous sections, what would happen if an AI agent could autonomously PR into main during a beta test? Is there any company comfortable with this level of AI autonomy? Security experts already warn of the principle of least privilege in regard to human developers and, while some companies shirk this responsibility in the name of progress, the industry will no doubt have to double down on this practice in an age where AI does work.
The unintentional oversights on behalf of the AI are only the start of our issues given excessive privileges. A few bumps and bruises on the security and stability sides are preferred over explicit agent hijacking, where the agentic systems are exploited directly or indirectly through prompt injection. An agent with keys to the kingdom and serviceably clever prompt injection are ingredients to a recipe which dumps confidential data, executes destructive commands, and worse.
Some businesses, in eagerness to leverage AI, may overlook this principle. Imagine: A company integrates the $10,000/month developer agent and, to make it useful, connects it to their most useful tools – code repositories, CI/CD systems, a task tracker, and maybe even the production environment directly – it is not hard to imagine the agent fetching secrets and exposing them or deleting the entire system if not provided proper security controls: The risk scales with the scope of an AI’s autonomy and the systems to which it has access.
In healthcare or finance, an AI agent’s mishandling would violate laws and patient privacy. In IoT and aviation, an agent misinterpretation of connected device controls would cause physical harm. Even in a mundane office setting, an agent can cause the fall of a business with the unintentional DDoS of a critical company site. At this point in time, individual and enterprise customers should pay attention. Who is making safe bets to not violate the value promise offered to you?
The prices for these AI agents raise clear red flags. While this article has covered the feasibility concerns of these agents from a high-level, I would be remiss not to cover the economic and ethical considerations of what seems a zero-sum game: one less human job and one more machine on payroll. Businesses may justify this by citing efficiency or talent shortages, but the broad implications of this trend leave some clearly disturbed.
A policy vacuum in this regime means businesses and workers largely navigate these changes alone, without the conveniences of safety nets or transition plans. The lack of a coordinated response or a forward-looking labor strategy is a stark omission given the scale of the potential upheaval: Highly educated workers pushed out of jobs en masse will result in increased underemployment across the middle class and unemployment across the lower and middle. With a smaller base of employed consumers and a loss of human capital development, career progression would be hollowed out with the so-called career ladder whittled to its upper rungs, a shift which shrinks the tax base and slows GDP growth over time.
Perhaps equally unfortunately, a heavy reliance on AI agents across national economies may stifle genuine innovation and give rise to those economies which either naturally emphasize AI less or lack the economic means to generally afford such agentic systems. With some psychologists reporting that AI over-reliance diminishes human creativity and innovation, an economy which holds generally steadfast on human labor, either through principle or need, seems the best optimized and time-tested approach for a national prosperity.
Without a clear path forward, each business with sufficient capital will evaluate for themselves the viability of paying six figures to rent an AI employee. Some professionals may answer with the following: If a $120,000/yr agent can replace more than one developer, it would be cost effective. However, there are positions for entry-level developers at $60,000/yr, and—unlike a human hire—the AI agent’s “experience” doesn’t deepen over time without additional investment from the company and OpenAI alike.
With the World Economic Forum suggesting “human-in-the-loop” oversight for AI agents to combat their risks, there are hidden costs to AI agents for those companies with stake in a quality output. The story of an attorney sanctioning ChatGPT to draft a brief, the agent conjuring and submitting fake cases, and the resulting court sanctions should serve as a stark reminder. Inevitably, however, businesses will fall into the same, producing substandard contracts and hoping to lose just a contract and not the business and allowing for more clever competitors to take the wheel.
As a Product Manager myself, I do not want to come in each day to greet a team of AI agents each morning at the Daily Scrum. Shirking the responsibility as an organization to provide jobs will produce a two-fold effect: The displacement of workers and the concentration of power into the hands of big tech firms. While we’ve seen and survived the impact of automation before, observing how technology-driven productivity boosts do not automatically translate into broad prosperity, the difference now is the magnitude and skill level of jobs on the chopping block. When the jobs at stake are those for which individuals have worked tirelessly, what will motivate them to continue? Some small solace may be found in identifying those organizations which inevitably fail for too aggressively replacing humans with AI, but the solace will be fleeting for those still searching for new employment. Of course, even if the AI agents mature and provide great value, there’s an ethical imperative to consider how we integrate them without discarding human workers. Success will be had for those companies which find the right balance, while those that cut people entirely may live only long enough to pay the price.