The weird world of working with AI agents
The weird world of working with AI agents
⏩ Future Normal: Fast Forward #110
The world of work is about to get…a little weird.
Weird, because you're about to start working with, and even managing, coworkers who aren't human. Yet what's even weirder? Before long, this will feel completely unremarkable. Normal, even.
Many of us are only just getting comfortable with talking to ChatGPT or Claude. And now Marc Benioff, CEO of Salesforce, is declaring AI agents as “The third wave of generative AI… what AI was meant to be.”
What's going on?
Sure, there’s a lot of media-friendly positioning here, but the promise of AI agents is compelling. The optimistic outcome is that they will invigorate our professional and personal lives—freeing us from routine tasks and enabling us to spend more time on creative and human-centric work. Of course, the alternative view sees us slip into a dystopian world of “agent says no” inhuman automation.
It’s still too early to know which future will become our normal.
But with all the big tech companies investing heavily in AI agents, it's essential that you start thinking about how they'll impact your organisation. To help, I'm sharing my working mental model – which outlines three potential ways we’ll experience AI agents:
1. Human-to-agent: Collaboration
3. Agent-to-human-to-human: Connection
Two quick caveats. These will inevitably blur together. After all, as the saying goes, “All models are wrong, but some are useful.” And, as always, I'm going to focus on the potential opportunities that AI agents could present you, rather than the risks.
Finally, while I've been working on this piece for a couple of weeks I'm still wrapping my head around all its implications.
So at the end I've added a list of 10 further questions about what this all might mean—for us as humans, for leaders, for organizations, and for society at large.
Let’s go…
A note from our sponsor (me!). You’re probably deep in sales kick offs, customer conferences, and 2025 strategic planning season. 🚨 I’d love to help.🚨
Typically I inspire clients with:
Your Future Normal: a trend keynote, bringing you relevant yet non-obvious insights from outside your industry.
VisuAIse Futures: a new interactive, ‘multiplayer’ AI-powered experience to unleash the creativity of your team. Watch the 2-min highlight video.
Read more about these at the bottom of this email, if you’re interested.
In the last few weeks I’ve had gigs in London, Bournemouth, Hamburg, Berlin, New York, and Las Vegas. Still to come: Rome, Greece, Barcelona and Dubai, as well as staying closer to home in London and Berkshire :)
I have a few slots left before the end of the year, and am taking bookings for 2025.
If you’d like to discuss bringing me to your next event, please do reach out to Renee Strom on renee@ideapress-speakers.com.
1. Human-to-agent: Collaboration
At Google I/O in May, Sundar Pichai shared a video of an AI agent autonomously scheduling the return of an unwanted pair of sneakers for a customer. While that was a consumer-facing concept, just a few months later it’s now clear that most people will get their first experience of these semi-autonomous, task-focused AI agents at work.
Here are a few early agents worth checking out:
Microsoft Copilot Agents. Microsoft unveiled the ‘Wave 2’ phase of its Copilot tool. As well as pushing Copilot into all its Office apps, it launched Copilot Agents. These can draw on business-specific documents and data sources to become more reliable and secure, and integrate with other systems in order to take action.
Salesforce Agentforce. Salesforce’s goal is to deploy 1 billion agents by the end of 2025 to “analyze data, make decisions, and take action on tasks like answering customer service inquiries, qualifying sales leads, and optimizing marketing campaigns.” Intercom’s Fin and Sierra are two other similar agents focused on automating customer service interactions.
Replit Agent. Coding was the first job to embrace copilots, so it’s unsurprising that it is also seeing the first AI agents. Cognition’s Devin was the first to wow people (read my earlier thoughts on how AI agents could make coding a ‘better’ job). Similarly, Replit’s coding agent enables users to create and deploy applications using just natural language prompts.1
Amazon Q Developer. Amazon’s agent is designed to accelerate software development, in areas such as testing, debugging, and finding security vulnerabilities. In August, CEO Andy Jassy shared that by automating the migration of applications to newer code, the tech company had saved 4,500 developer years (!!) of work and unlocked annual cost savings of $260 million.
Paradigm. As someone who spent the first 5,000+ hours of my career building financial models in Excel (my dirty career secret is that I’m a trained accountant!) I’m excited for this new startup which aims to make spreadsheets radically simple. No formulas. No macros. No out-of-date data. Its AI agents simply get the data you ask for. I geeked out hard on its demo video (below).
2. Agent-to-agent: Delegation
Most readers I’m guessing are pretty much on board with the agents above. Who wouldn’t want a swarm of AI agents taking care of many of your routine tasks?
Now let’s turn the weirdness dial up a bit.
Because you don’t have to stretch your imagination too far to see that the next step will be agents interacting with other agents. Right now, this is mainly confined to research demos and games. But it’s clear that it could truly disrupt our notion of what ‘work’ looks like.
It can also quickly get dystopian. Or comical. When I learned that companies like Fairgo have created AI recruitment agents that can conduct initial screening interviews, my second thought (after the initial WTF moment!) was that surely candidates will just send their AI avatars to these calls, too?
More broadly, agent-to-agent interactions promise to enable harder, multi-stage tasks to be completed, and unlock more complex social simulations:
Altera’s Project Sid. The video of this startup’s 1,000 AI-generated characters playing Minecraft went viral recently (see below, it’s well worth watching!). The company explains its mission: “as Richard Feynman said, “What I cannot create, I do not understand.” Creating digital humans is our way of understanding humanity better, both at the individual and societal levels.”
Crew AI. This platform enables users to build multi-agent systems that can work together to automate complex tasks. As IBM reports, “multi-agent systems (MAS) divvy up tasks among several specialized agents … [and] demonstrate notable advantages over chain-of-thought prompting, where the model needs to break down tasks into a series of steps. Multi-agent architectures tend to thrive more when collaboration and multiple distinct execution paths are required.”
Coinbase’s AI-to-AI transaction. Brian Armstrong, Coinbase’s CEO, said that these transactions are “an important step to AIs getting useful work done. [Today] AIs can't transact to acquire the resources they need. They don't have a credit card to use AWS, Github, or Vercel. They don't have a payment method to book you the plane ticket or hotel for your upcoming trip. They can't get through paywalls (for instance to read a scientific article), or promote their post on X with a paid ad.”
3. Agent-to-human-to-human: Connection
However, most discussion about agents focuses on their capability to perform known tasks (including this newsletter up to this point!). It lacks imagination—and humanity.
Contrast this with Seth Godin’s recent piece, “A Possible AI Future”. It’s so good that it’s worth quoting in full here:
Most visions of the internet in 1995 were about individuals interacting with content online. It turns out that the internet is actually about connection. The apps and businesses that were most successful connected people–to ideas, to things or mostly, to each other.
The current range of AI feels like content creation. You can have an AI write your high school essay, draw you a picture or invent a recipe. But perhaps history will repeat itself. Perhaps developers will realize that persistent knowledge of what came before and who needs help and connection is the next frontier.
So… what happens when the AI in our lives begins acting like a thoughtful, patient and trusted friend? Not just like the AI in the movie Her, but more focused on our networks and connections. Who’s trustworthy, talented, available or in need… It knows what’s happening now, but also what happened yesterday. Not just to us, but to those in our circle and the people they know as well…
You’re about to throw out an old board game from the attic. The AI whispers, “Hold a sec, I think a neighbor down the street has been looking for something just like that–want me to sell it to him?”
A company seeking RFPs invites all its suppliers to submit confidential overviews of their supply chain. An AI reads the material and creates Pareto optimal connections, building a confederation of several suppliers who can work together to build something faster and more efficiently than any could do alone.
Three people are leaving a conference and they’re all calling a car to take them to the airport. Perhaps the AI offers a carpool.
We’re headed off to a community meeting, and we let the AI know that if someone there is hiring for the kind of job we’re good at, we’re open to a connection.
This is a level of intimacy, attention sharing and data that dwarfs anything that has come before, and it brings with it huge issues of permission, control and privacy.
I can think of a thousand ways that this power could be misused, manipulated and go terribly wrong. I have also seen the internet go wrong too.
But this is only the beginning of the AI age, and it might help to find a north star, a standard for what happens when the connection machine works for us, instead of against us.
It’s an ambitious vision. A positive vision. A deeply human vision.
Indeed its ambition makes me wonder – what if the natural starting point for this kind agentic layer was within an organisation?
Companies have huge incentives to use AI agents to break down internal silos, and to transform internal collaboration from the chance activity it so often is today (see Steve Jobs’ design of the Pixar office), to something much more intelligent and strategic.
It would still be a huge challenge, but the payoffs – for both the individuals involved and the organisation – would be much more tangible than the broader, public agents that Godin imagines above.
Right now this remains still largely conceptual, although startups like Fractal are exploring what the company ‘operating system’ in the agentic era could look like. It’s an area that I’m watching closely. And so should you ;)
Where next? 10 further, non-obvious questions about our agentic future.
In the intro to this piece, I said this is still very much a working model, which opens up as many questions as it answers. Here are 10 swirling around in my head:
What’s the best analogy for these AI coworkers? Which term will win out? Agents? Copilots? Assistants? Coaches? Or will we all have swarms of various agents? Personally, I still love Descript’s ‘Underlord’.
How will we ensure we can trust AI actors’ output? It’s a big step to delegate an end-to-end action to an autonomous agent. Yet we will, eventually. Although paradoxically as AI agents become more accurate, the risk of people ‘falling asleep at the wheel’ and missing infrequent and/or subtle errors becomes even greater. Whatever happens, I suspect it will be a great time to be a lawyer, as we wrestle with issues like IP, accountability and more ;)
Will we need to ‘add an egg’ into the AI agent mix? Rex Woodberry makes a great analogy between AI agents and instant cake mixes. Initially, cake mix sales were slow, because it felt too easy. They only took off once you had to add an egg – people wanted to feel like they had contributed.
Where will we find meaning in our work? Will it lead to an (even bigger) crisis of meaning if many desk-based jobs become even more about coordination than doing the actual work?
How will agents rejig the status game? Will having a swarm of AI agents acting on your behalf be a status symbol? Similar to having a Blackberry in the pre-smartphone era? Or will delegating tasks to real, flesh-and-blood humans be the ultimate power flex, similar to the ability to take an extended digital detox is today?
What skills will humans need? I always think of Scott Belsky’s brilliant quote, “in the age of AI, taste will become more important than skills.” Which of course leads to the bigger question – how will we teach great taste?
What will it mean for leadership, if people start to trust AI agents more than people? When agents can successfully complete many tasks, then will setting the right goal – asking the right question – become the determinant of a good leader?
How will agents disrupt business models? Most agencies (who employ humans) still charge for time spent. Most online services pay for seats used. None of these make any sense in a world of AI agents. Intercom’s and Salesforce’s customer service agents both charge per successful resolution. One challenge I can foresee – our expectations of ‘success’ change fast.
Who will benefit most from AI agents? Will agents enable small teams to create huge businesses? Or will bigger companies with vast amounts of data be able to deploy better, more effective agents? On an individual level, will AI agents reduce inequality by reducing the cost of labour? Or will ‘premium’ AI agents exacerbate it?
Will AI agents be able to create authentic human connections? If my AI agent negotiates with your AI agent for us to meet (in real life!), will we view this as somehow less ‘authentic’? Perhaps initially. Although I suspect that just as with online dating, if it works then any stigma will be short-lived.
I don’t have answers to most, if any, of these questions. But with as we approach the agentic era, you can’t afford not to be asking them.
Can I inspire your team to seize the future?
This year I’ve delivered 25+ sessions, both live and virtually – from Brazil to Saudi Arabia, Slovenia to Shoreditch.
As well as my regular trend & innovation keynotes, I’m hugely excited about the reactions to my newest offering – VisuAIse Futures.
It’s an interactive, ‘multiplayer’ creative experience that will leave your audience thinking differently about AI:
“It was so refreshing to hear how AI can be used to power human imagination, rather than replace it. And then it was even better to actually experience it”
“Fantastic session! Hugely insightful and fun, too!”
“Brilliant. The feeling in the room was positively intense whilst the images were coming through!
Feel the optimistic vibes it will bring to your event in the 2-minute video below.
If you’d like to discuss bringing me to your next meeting or event then please do reach out directly to Renee Strom or check out my speaking site.
Thanks for reading,
Henry
One big concern about AI-generated code is that it may contain security vulnerabilities. But of course, there’s now an AI agent for that.