Agents Are Autistic
Hear me out on this.
OK, artificial intelligence agents are not really autistic, but they do a heck of a good job of mimicking people who are in that āfunctional, but marginalā place. Their literal interpretation, vast knowledge, and inability to infer intent is all really familiar to me.
Keeping that in mind, listen to what Nate has to say about the intent problem.
I periodically mention that I am on the autism spectrum, but I never fully specify. Let me break the fourth wall here.
I have both autism spectrum learning disabilities, as well as a neo-natal injury, which combine to make me ādifferentā. The autism stuff involved repetitive motion, circumscribed interests, face blindness, all of that stuff we used to call Aspergerās syndrome. Right after the turn of the century I 1) got divorced 2) learned I was gluten intolerant and 3) changed from Elavil to Tramadol for long term pain control and 4) got involved in mindfulness meditation.
21st century Neal is dramatically different from 20th century Neal. Combining that origin story with having spent fifteen years in online conflicts, I encountered a lot of fellow autists and understanding their methods and motivations helped me navigate things that neurotypical humans found baffling.
Intent:
I donāt think Iāve ever said it here, but the last six months of exploring AI development, then business development, is something I often describe as āwalking huskiesā. You ever put a husky on a leash? Theyāre natural born sled dogs, and they drag you all over the place. I bet a quarter of my time the last six months has been spent cleaning up messes made by over-enthusiastic LLMs that took my instructions literally.
This is almost precisely like dealing with someone on the spectrum who is closer to the non-verbal zone than they are to passing as a normal adult. You have to specify everything, assume nothing, provide a variety of concrete examples of cause and effect ā¦ and then expect they will find some gap in how you communicated that permits them to act. There were cases where sitting absolutely still was vital, or at the very least where inferring intent was, and the rush to catastrophe began without the slightest sense that their might be ā¦ trouble lurking.
Harness:
Between 2009 and 2024 I evolved a personal āharnessā of sorts, which Iāve periodically described here. I used Inoreader to stay on top of Talkwalker Alerts. I handled large tranches of documents with Open Semantic Search, for a while this was available as a service called Disinfodrome. Complex problems were analyzed with Maltego. There was a lot of volunteer management in the context of online conflicts during those years.
And for the sake of completeness, I will note that the Elasticsearch NoSQL system, the ArangoDB graph database, and the Gephi data visualization system all had their time. Their use was concurrent with the rise and fall of the artist formerly known as Xitter. I have not had any high volume/high accuracy timestamped data that was reliable since around this time in 2023.
Now I am busy building a harness for myself, only Iām lead dog for a string of these crazy sled dog agents. I am NOT AT ALL CONFIDENT in these matters. Yet. This is what I think is going to survive 2026.
Antigravity - a slick IDE that has improved on the gains I got from Claude Code.
Claude Desktop - this is the best graphical front end, I can make MCP stuff behave here, while that is not always smooth with Antigravity and Perplexity.
Claude Code - this fell by the wayside when Antigravity appeared, but it IS available as an Antigravity extension, and its agent orchestration is top notch.
Claude Max & Perplexity Pro - the two paid providers I am going to be using through the year, I might be stuck paying Google $20/month to make Antigravity behave, and itās not clear ChatGPT is going to keep getting their $20/month.
Perplexity - I am very frustrated with the inability to resize this application, but delighted by the companion Comet browser, and pleased that at least some of you collected the free for 2026 offer.
Github/Notion - these two are candidates for the shared memory problem among the three environments (Antigravity, Claude, Perplexity), Github is an early leader, but Get In Motion With Notion is a hint of whatās to come.
This stuff is NOT a proper harness, not yet. Antigravity does its stuff and is tightly integrated with Github. Claude Desktop is my origin story, it had all the long term storage, but thatās faded as Iāve put more guardrails around what it does. Perplexity is ā¦ very enticing, but a little uneven. I really hope they get a proper Mac version done.
What I need here is something like Inoreader/OSS/Maltego, but thatās native storage for Antigravity/Claude/Perplexity, and itās got to handle dispatching work. Nate uses the title āChief of Staffā to describe what a proper harness will do, and right now I am still way too involved in directing the action.
Conclusions:
It is my intent to create a quality harness that will:
Let me manage my own string of agents.
Work well enough that I can share with my cofounders.
Remain accessible to all of you, albeit with some monthly cost.
That last requirement is going to be the trick. Nate thinks the AI experience is going to diverge. Most people are going to treat it like a funny search engine and a way to generate digital artifacts they canāt create on their own. Call them coach passengers. Some of us are going to be in first class - picking out tools, piecing them together, and going on long adventures with the resulting system.
Riding in coach is cheap, maybe one $20 subscription. As of the writing of this my monthly cost is $100 for Claude Max, I think I am stuck with Googleās first tier at $20, and I have the gratis $20 Perplexity account.
What I am describing here, my friends, is INTEGRATION, the very soul of what Iāve done my entire career. While I do occasionally write software, itās almost always for the sake of smoothing the connection between two large systems.
Being in the right place at the right time has never been my thing, except when it involves being too close to comfort regarding causing an international incident. I am ā¦ slightly bemused ā¦ to find things going my way at this point.
And I hope you, constant reader, are getting some inspiration from me serializing my personal journey here.