Awash in agents

Brad Silverberg reminded me that this past weekend was the 30th anniversary of the release-to-manufacturing of Windows 95.  Back in those pre-Internet days, we had to hand off final bits to our manufacturing plant to create CD and floppy disk SKUs, and we had to hand off final bits to our OEM partners for them to install on new PCs.   It was a scary day — updating software in the event of a bug was a LOT harder then — and it was a thrilling day, the end of an incredible product push with a great team. 

Like many of Microsoft’s products, Windows 95 was driven by, and was a bet on, Moore’s Law.  We all knew that computational power would continue to expand dramatically, and that we needed to build software to use that power and make products easier, more productive, and more enjoyable.  Windows 95 was a big step up in computational demand — 32-bit, GUI, networking, DirectX games support, etc.   And throughout Microsoft, other bets were being placed on software that would use even more compute power — apps, Windows NT, and so on.  We had to drive to keep up with Moore’s law constantly.

Betting on Moore’s law was the right bet then.  It’s been the right bet for 40 years.  It still is the right bet.

AI is going to get way cheaper

It is early days for AI.  As Laudon Williams says, “Today, right now this minute, AI is the worst it will ever be.” Inference costs will continue to drop dramatically, allowing a dramatic increase in use.   We are going to use way more AI, and it is going to get way better.

Cheap, abundant agents will create whole new opportunities, all using a dramatic amount of AI capability.   We will want to figure out how to:

• Spin up tons of agents.  As inference costs fall, the barrier to spinning up persistent, specialized agents disappears.  We need “Heroku for agents” — platforms that let us spin up agents effortlessly and cheaply.
• Run swarms of agents.   With cheap inference, we can easily spin up a whole team of agents to work on our behalf.  We need tools to coordinate and direct our teams of agents easily.
• Rewrite current applications as agents.  We can replace the static brittle app architectures of today with flexible agents.  Agent-native blogs and newsletters, agent-native wikis and KBs (Confluence is still a heavily used corporate app — seems ripe for replacement), agent-native CRMs, etc.  These are all just databases with a bunch of static view code and static business logic; they are ripe for replacement with an agent generating dynamic views and dynamic inferences across their datasets.
• Run an entire business with agents.  Agents create all plans, handle all regulatory and incorporation filings, create landing pages and prototypes, track engagement, and create a business pipeline — all from a single mission statement.
• Create always-on persistent agents.  For instance, your personal career coach which looks at your calendar, all your correspondence, all your contacts, all your web activity, and suggests how to focus time, how to prep for meetings, and how to manage your pipeline of activities.
• Manage all our agent spending.   Agents and inference are going to be the single largest IT expense item for most companies and people.

And probably 100 other things.  Anyway, this is where the fun is — figuring out how to yoke together a large amount of AI/agent/inference technology into coherent and effective solutions.

In my own little newsletter effort, I am struggling daily with how to balance all my agent activity, where my budget should be, how to sequence and prioritize tasks, and how to effectively coordinate agent efforts – which is as it should be.  An agent-based newsletter effort should probably consist of a team of agents, all working together:

• A production agent. Tracks my writing pace, goads me to write, manages all my snippets, and suggests how to relate them to current events.  Watches the news and other newsletters and suggests topics
• A writing agent.  Captures ideas, suggests room for expansion, critiques what I write.  But never actually writes posts!
• A publishing agent. Publishes the newsletter, provides snippets for social posting, etc.
• An engagement agent. Watches discussions and social feedback and suggests topics for response and elaboration

Currently, I handle all those tasks, but I want to delegate them to agents as soon as possible.

Experiments in multi-agent use

I mentioned last week using Vibe-kanban and Claude Flow to manage agent activity.  I still like Vibe-kanban a lot.

Rohit Krishnan has pushed this idea ahead a lot — his concept of a dashboard to manage all his agent activity appeals to me.     I want something like this!   I may have to attempt it this week.

Ethan Mollick mentioned the ability of ChatGPT5 to perform regularly scheduled tasks, and I have been trying this out as well.  It doesn’t seem quite ready for primetime yet — as seen here, it failed to complete my first scheduled task for completely unknown reasons, and there is no detail available at the site.  But a very worthy idea.

I am enthused about experimenting with any tool that allows me to spin up and manage multiple agents – if you have any favorites, let me know.