"We're All Executives Now": How 'Agents' could change how we work in '25

I continue to be impressed by Cognition 's Devin, the first legitimately ~autonomous "AI software engineer": It's forcing me to level up (IC to manager mindset/skillset when it comes to product development for my AI Chief-of-Staff product [now used by 9 CEOs across the US!], which I've been mostly brute forcing with verbal direction + o1-preview and now o1 Pro), which is a nice bonus, beyond the core benefit of getting time back.

I'm not alone in this assessment; there are even reports of teams pulling software engineering job listings after just one week of use:

It's also spurred a few ideas about where things could/might/should go if we extend the general "agent" approach (if you want to go deeper, I'd highly recommend this salient article from the team at Anthropic ), spurred by the following question that's implicitly raised by these early signs of life from D., o1, et cetera:

"What are types of work (and maybe, more specifically, precise slices of soon-to-be-vestigial knowledge work roles) where:

• the agent can be narrowly-enough spec'd in re it's general goal and specialized toolset, such that
• simply cranking up compute spend would
• yield a result that is at least 50% as good at 1/10 the cost of the human labor substitute, and perhaps
• unlocks previously inaccessible value
• due to the enlarged scope of the information space that can be searched and mined due to
• the lower cost and ~10x faster computation?"

In general, I think the answer is "research"; there are narrow problem spaces where one could build a profitable, tailored research assistant product that is:

1. more bespoke than, e.g., Operator, Mariner, and whatever Anthropic releases;
2. has access to semi-proprietary data sources—or, even better, an expanding "data commons"; and
3. commands a nice premium over traditional SaaS seat-based licenses by trading a massive reduction in marginal clicks for high-3-to-even-4-figures of revenue per month (particularly relevant: the perennial Stratechery insight in re the relative impossibility of "overshooting" good UX).

Three possible instantiations of the above (yes, this is an explicit invitation to build these—I would pay a premium for all three!):

1. a "business developer" that facilitates fruitful connections via deep, deep research prior to bidding for the principal's limited attention (related: one of the main bottlenecks going forward will be human attention, i.e., the rate of improvement of context window length is surely greater than our own, or we're all executives now); an MVP of this might be a matchmaking system that facilitates the fewest, most surprising connections between conference attendees prior to the event given some publicly available information about the ticketholders and a modicum of information in re their goals, what's salient, etc., all of which can be elicited during a short phone conversation via a voice assistant like Vapi ;
2. a "recruiter" that does search across not-only-human candidates and even breaks the FTE/contractor models by default: Given some rich data (again, let's say, meeting or call transcripts), this flavor of agentic system should be able to derive the no. 1 problem their principal faces (I can tell you from experience, by the way, that this is often latent in the unstructured source material, not apparent those who face the problem, but legible to state-of-the-art foundation models), performs a massively broad search across specialist software solutions (which could be products for people or tools for agents with deeper access to their principal's secrets), hyper-differentiated individual SMEs (i.e., AI-powered services solopreneurs, which is already emerging labor category...I know this, in part, because I am one) who are open to outcome-based compensation, and then, on the margin, your more traditional specialist shops/agencies or—GASP—a part- or full-time hire; and
3. an "aligned curator" that produces some kind of anti-mimetic content feed: Instead of hoovering up attention by prioritizing socially proven content that hooks directly into our lizard brains à la the web2 behemoths, what if your open-ended discovery agent (see also: Ken Stanley's work) could go deeper and deeper into an untrodden pocket of latent space, generating hypotheses and even counterintuitive insights for you to meditate on, say, a weekly basis? I'm talking actual personalization (in the sense of personalism and studiositas vs. the potent combination of pride/envy/curiositas that drives so much digital media consumption at the moment).

Perplexity , Google's new Gemini Deep Research product, and o1 with tools and variable TTC are great first steps.

I look forward to seeing how entrepreneurs, including and especially YOU!, take the aforementioned capabilities to the next level in '25—and reporting back on what I learn as I integrate these nascent, next-gen AI capabilities deeper into my and my clients' workflows, all the way from single-player to multiplayer modes.