Agents at Work
I’ve been spending (arguably) too much time with AI, trying to put the technology to good use across the different domains of my life: family and personal management, standing up a new VC firm called TK Ventures, and the ideas conference I co-founded called Dent.
I now divide my AI use into two categories: AI assistance while I’m doing something, and AI-powered workflows that do something on my behalf.
It’s become clear to me that the most useful workflows are going to be ones that can be set to run independently and somewhere either in the cloud or on a machine that is always on and connected. Some examples:
I transcribe many of my zoom meetings. Every evening, these transcripts are reviewed for tasks and follow ups, compared against the tasks already in Monday.com, and new ones are added to my backlog. I then get an email summary of the changes.
Every day, an agent reads the last 24 hours of newsletters I’ve received and, based on my Substack posts and an extensive style guide, drafts a custom blog post idea based on a news hook from what I’ve received. (I haven’t posted any of these, but some of them are quite good and I’m realizing it might be a better way to read my newsletters than trying to slog through all of them).
Every week an agent reviews my calendar and my email for changes to my travel schedule, compares them to my wife’s travel schedule, flags conflicts or overlaps, and sends me an email. Then it scrubs the personal stuff and sends my travel calendar to my partners at TK and at Dent.
For now, these all run on my computer and use either APIs or MCP connections to manipulate the external services involved. But I would really like it if these and other workflows could run on independent hardware.
So far this is a challenge because there are real security concerns with giving AI tools access to your accounts. The consensus seems to be that AI agents are best considered “employees” for security and provisioning purposes: each agent should get their own accounts and those accounts should have shared access to only the parts they need.
My partner Stewart sent me this great summary from Union Square Ventures describing how they’ve been setting up agents to make their workflows better. It is a great blueprint and I’m really glad they’re sharing it publicly. As a result, I spent time today using Tasklet to connect our TK meeting notes with our dealflow pipeline in Attio, and the setup experience was far better than doing the same in OpenClaw or Claude Code.
It remains to be seen if the workflow will actually run when it is supposed to run without human intervention. In my experience most of these workflows run great when triggered manually and don’t yet run well or don’t run at all when they’re supposed to run by themselves.
I need to write a whole different post on how Claude has become my default interface to SaaS services. I have always hated 90% of the interfaces I have been forced to use online, and the fact that Claude can coordinate between multiple services using multiple data sources has been an amazing improvement.
But all of this is exciting. Tasklet seems to have leapfrogged Make.com and Zapier by taking a page from OpenClaw. There are still some very obvious (but difficult) gaps driven by what people want to do to make LLMs useful and it’s going to be very interesting to see which companies manage to take the lead in filling them.