I.

Is Anyone Working on Agentic Authentication?

By In Technology

Everyone is building AI-powered tools, even people who shouldn’t be. Agents seem to be the next obvious (and big?) step. But these little bots need a secure way to act on behalf of users without causing chaos.

Richard Dulude at Underscore VC wrote about the lack of identity standards for AI agents in this LinkedIn article. I don’t know Richard or Underscore VC (sorry). But, he’s right, traditional authentication assumes either a human or a machine with static credentials, and that doesn’t work for AI agents that need to make decisions and take actions. Companies want accountability (and probably liability), and users need control of what their potentially psychedelic robot is doing on their behalf. This balance doesn’t exist yet.

This is probably for another blog post, but right now, everyone, including the bots, are using human interfaces as a stopgap. OpenAI’s Operator is a great example, agents pretending to be humans to interact with systems that weren’t built for them. That’s fine for now, but eventually, the human interfaces will be an afterthought. Like how “mobile-first” design took over, we’ll be doing “agent-first” design with human-accessible backups. Having a dedicated standard for agentic authentication might be a good first step in that machine-to-machine way of thinking and designing systems.

Agentic Proxy Credentials (APC): A Solution (A Term I Totally Made Up)

I made this up. It’s probably a bad term, but naming things is fun. This doesn’t exist… if you are a large battery and power supply company, don’t sue me. I’m spitballing here.

One possible fix is the “sucked out of my thumb” Agentic Proxy Credentials (APC). This would let users grant their AI agents secure, limited permissions to interact with systems while making sure the right level of oversight are in place. There are things that I wanted to do this very week, but I don’t trust my bots with my actual usernames and passwords:

Stop me talking to Airline Idiot Bots

Talking to airline chatbots is painful. Right now, they can only regurgitate FAQ answers. With an APC, my AI assistant could log into my airline account, check flights based on my loyalty status, and rebook me without you having to touch anything. This would make AI actually useful instead of just a slightly smarter help page.

Paying for small things without having to deal with entering my ACH data AGAIN

I don’t want to give an AI full access to my bank account. But I wouldn’t mind letting it handle small transactions in a controlled way. With APCs, I could grant my assistant time-limited access to approve payments or move money within strict limits. The AI does the work, I stay in control, and my bank account doesn’t mysteriously empty overnight… unless I’m Ambien shopping again.

AI Dungeon Master’s Assistant

D&D is great, but session prep is a time sink. I want an AI that logs into my D&D Beyond account, manages stat blocks, generates lore-friendly content, and even takes session notes. The AI handles the boring admin work, and you get to focus on making your players cry (or cheer, if you’re nice). Yes, serious stuff here.

How It Could Work

There are a few ways to make this happen, I think. I’m no longer allowed to do actual engineering at my own companies I founded, so this blog is my outlet. Everyone needs a hobby.

Is Someone Already Building This?

Honestly, I wouldn’t be surprised if Okta, OAuth, or OpenAI are already working on this and I’m just ranting for no reason. But if they aren’t, they should be. The pieces are all there, someone just has to put them together.

I need this, but I can’t find it. If anyone is working on it, let me know. I’m too busy trying to solve employee gifting at scale at Thankscrate, implementing AI into every existing business at Sourcetoad, and making sure passengers can watch TV and book dinner reservations in the middle of nowhere at OnDeck.

Previous ArticleNext Article
I help companies turn their technical ideas into reality.

CEO @Sourcetoad and @OnDeck

Founder of Thankscrate and Data and Sons

Author of Herding Cats and Coders

Fan of judo, squash, whiskey, aggressive inline, and temperamental British sports cars.

Leave a Reply

Image generation comparison from February 2026

In Technology

I spend a lot of time generating images these days for presentations. My typical workflow is fairly scientific: I ask Midjourney to produce a relatively cute image of a frog, a toad, a robot, or some other vaguely anthropomorphic creature doing something related to the slide I’m about to present.

Once I get the image, I expand the background by about 90% so the character ends up in the corner of the slide. That gives me a nice, relatively clean area to drop text on top. Sometimes I use Photoshop to do the expansion. Sometimes Midjourney cooperates. ChatGPT is actually pretty good at this too. Nano Banana is… enthusiastic. It tends to try a little too hard right now.

That’s fun and all. But the more interesting comparison isn’t cute amphibians. It’s boring enterprise diagrams.

Recently I had to generate some architecture visuals for an RFP response. Rather than suffer alone, I decided to turn it into a model comparison experiment.

Below is a slightly simplified (but very real-feeling) prompt I used. The company is fictional. The buzzwords are not:

Create a clean, executive-level architecture diagram titled “Closed-Loop Member Intelligence Platform.”

The layout should be 16:9 and structured left to right with a circular optimization loop surrounding the system.

On the left side, show multiple member touchpoints feeding into the platform:
- Website (class browsing, account login)
- Mobile App (workout tracking, push notifications)
- In-Club Kiosks (check-in terminals)
- Wearable Device Integrations (fitness trackers)

Label this section: “Member Interactions Across Digital & Physical Channels.”

All touchpoints should flow into a large central hub labeled:

“Unified Member Profile & Real-Time Event Engine”

Inside the central hub, include:

- Web SDK
- Mobile SDK
- API Gateway
- Event Streaming Layer
- Clickstream Data Capture
- CRM Data Sync
- Identity Resolution Engine

Include a small sub-caption:
“Event-level data unifies anonymous visitors and active members into a single dynamic profile.”

From the central hub, arrows should flow to a right-side activation layer labeled:

“Real-Time Engagement & Orchestration”

Include these outputs:

- Personalized Workout Recommendations
- Dynamic Class Availability Messaging
- Triggered Retention Offers
- Membership Upgrade Campaigns
- A/B Testing & Experimentation Engine

Surround the entire diagram with a circular arrow labeled:

“Continuous Optimization & Revenue Growth”

Along the circular loop, include metrics:

- Engagement
- Conversion
- Retention
- Lifetime Value

Design style should be modern, minimal, and suitable for an enterprise SaaS presentation.
Use neutral tones with one accent color to indicate data flow.
Avoid clutter.
Make the architecture clear and readable for both technical and executive audiences.

Here are the results.

ChatGPT

Clear winner for “looks like a human consultant made this at 11:30 p.m. before a board meeting.” The text was incredibly legible. The layout was balanced. The hierarchy made sense. It genuinely looked like something you’d expect in a mid-market SaaS pitch deck.

I even did a reverse image search on some of the icons. No exact matches. That suggests they were generated rather than assembled from some common icon pack. Which is pretty cool.

Claude

Claude did something interesting. Instead of just giving me a static diagram, it generated a React application that rendered the architecture visually inside its canvas. I should have guessed this is what that nerd would do… in fact I did guess, but whatever.

That has upsides. I can tweak the code. I can modify the layout. I can version control it. That’s appealing to the nerd in me.

But technically it failed the homework assignent. It wasn’t what I asked for. I asked for a diagram image. What I got was a React app that displayed a diagram that I had to screenshot.

That said, I actually liked the aesthetic. It felt a little more “me.” Slightly less textbook. Slightly more product-thinking.

Gemini (Nano Banana)

The undisputed champion of 2026 in image generation, nano banana, was actually my least favorite of all of the designs. I think there’s something really weird about the arrows on the outside ring of this diagram. Why are there two arrows between “Engagement” and “Conversion”? Why are they different sizes? I did actually find a couple of exact matches when searching for some of these icons here, so so there might be some assembly on top of generation going here, but I cannot tell because these icons are so universal that it’s likely that that could just be a coincidence.

Midjourney

Ah, Midjourney. My current favorite for keynote frogs.

Completely and utterly useless for generating readable diagrams.

It’s phenomenal at stylized imagery. I’ve tuned it so much over time that it practically knows my aesthetic preferences better than I do. It’s like it’s been trained specifically to make amphibians that align with my personality.

The Omni feature (object permanence) is genuinely impressive. If you’re telling a visual story and need a character to look consistent across multiple scenes, or you’re creating a children’s book to convince your six-year-old that haircuts are not a violation of human rights, it’s fantastic.

But enterprise architecture diagrams? Nope, sucksville.

Wrapping Up

I was pretty sure that nano banana was going to run away with this one. Everyone I know works in banking or finance or medicine has been telling me how great the model is for generating diagrams and process flows. They’ve been raving about how things that were not possible three months ago are now completely durable with this model. It was a little bit of a surprise to see that my personal favorite was good old-fashioned ChatGPT. I think, for my personal use, I’m probably going to use Claude to generate diagrams because they’re a lot easier for me to tweak once they’ve been generated.

That said, I think this experiment showed that when I do this kind of work in the future, I’m just going to load up the same prompt in three different models and just pick the one I like the most. Some of it’s going to be personal tastes; some of it’s going to be how well the model interpreted the prompt, and some of it’s going to be the state of that particular LLM and its model on that given week.

And I’m going to stick to only using Midjourney for generating cute pictures of toads.