It’s humorous. On the one hand, I might say I’m a visible particular person in that I reply nicely to photos when making an attempt to be taught and keep in mind issues. Then again, I don’t simply retain these photos. My recollections and my method by which I navigate my historic life is extra influenced by story, metaphor and emotions than it’s by these pictures themselves.
Earlier than AI Brokers, like Claude Code, I now discover that my model of workflow had very a lot turn out to be centered round a hierarchy of folders, rigorously separated by subjects. I, like many individuals, got here to try this, however for me it was extra out of necessity than the pure emotions that it was required. I think that it’s because my thoughts is just not essentially the most organized. I’m not a messy particular person, however I wouldn’t say I do know the logical option to set up my residence, as an example. Ought to the sofa go right here, or ought to it go there? Do these colours work with the daylight? What about this plant or crops? The place ought to I put the mattress?
For me these could as nicely be making an attempt to determine convey what a skyscraper seems to be like utilizing nothing however interpretative dance. I merely can not simply expertise nor absolutely comprehend no matter it requires to reply questions round optimum group, and so I’ve usually been pressured to invent my very own group in response to catastrophic failures. Like discovering coding errors early on in my profession, which led me to analysis how others set up their folders, identify their information, and automate this, that and the opposite.
However the motion into utilizing extensively AI brokers for my analysis has created new issues for me. My perspective round AI brokers usually is that they clear up lots of my historic issues but in addition create new ones. I don’t lament this; I simply see it as my job to unravel these issues. I maintain no sympathy for the dying world, anymore than I really feel sympathy for misplaced days of typewriters, buggies and poorly match boots.
And but a few of these issues are like the outline above on the subject of arranging the furnishings in my home. The association of furnishings in the home is endogenous. It’s endogenous to the format of the home, the place the pure gentle is available in, the place supporting partitions are, how giant the home windows are, and maybe greater than all, what experiences I need to have, phenomenologically, when in that room. However it is usually endogenous to my tastes. And I can not clear up this endogeneity downside with out talent and human capital in one thing as fundamental as design.
Effectively, I feel it’s the identical with utilizing AI brokers to be trustworthy. The construction of labor has fully shifted, however not like the fashions I research in economics, I do know that the creation of the brand new equilibrium would require that I personally create one thing new.
I feel up to now I’ve been targeted on /abilities, virtually completely. Virtually atomistic like. Do my regular factor, however use a /talent to do it. So preserve making decks, however use /beautiful_deck to make the decks. Maintain writing code, however use /referee2 to examine and validate the code. However one thing is just not working completely, and I can not but fairly put my finger on what it’s, why it’s, and arrange the brand new workflow.
I’m going to reread David Allen’s basic productiveness e-book Getting Issues Finished, or at the very least begin there, to try to see if I can not perceive from core rules redesign my workflow. Similar to I had as soon as made the structuring of folders my core foundational blueprint to analysis, I must go even deeper. I want to determine create the optimum human-in-the-loop system. The place do I insert myself precisely on this new manufacturing operate? Recall from this quote from William Shockley, the 1957 winner of the Nobel Prize in physics, whereby he laid out a easy conjecture that the manufacturing operate of scientific output adopted a multiplicative mannequin that led to a log regular distribution in productiveness in scientists.
I’m intent on leveraging and squeezing each final drop out of AI brokers to reinforce that whole pipeline for me. However to try this, I’ve to begin over from fundamentals and take into consideration what’s the workflow of analysis in a manner that is so simple as humanly attainable. I’ve to have a way, as an example, to file my conjectures and interpretations given I’m not extra levels faraway from the typing of precise code. Such typing duties had traditionally, for some causes, positioned me right into a form of circulate state the place concepts bought imprinted, almost completely, right into a form of muscle reminiscence. So now the place do I put them?
And possibly whilst importantly — how will I visualize these concepts? Will I solely depend on “lovely decks”? May I create dashboards? Do I make markdowns of each thought? One markdown? Fifty markdowns? In a folder? How a lot do I hand over? How a lot do I defend as my very own territory? What are the property rights between me and Claude Code precisely, over what duties, and why?
So, I feel that my subsequent objective is not going to be a /talent. My subsequent objective is the AI Agent equal of developing with a brand new workflow, and that is in all probability the place I’m going to have to begin creating harnesses. What’s an AI Agent harness? What’s it particularly for financial and social scientific analysis? Think about this definition.
AI brokers at present are extra than simply standalone fashions that absorb and output textual content tokens. They function inside an ecosystem of instruments, reminiscence shops, and orchestrated workflows that allow them to carry out complicated duties. On this context, a brand new time period has emerged within the AI lexicon: the “harness.”
In easy phrases, an agent harness is the software program infrastructure that wraps round a big language mannequin (LLM) or AI agent, dealing with all the things besides the mannequin itself. One AI architect defines an agent harness as “the whole architectural system surrounding an LLM that manages the lifecycle of context: from intent seize via specification, compilation, execution, verification, and persistence”, basically all the things besides the LLM itself. In sensible phrases, the harness is what connects an AI mannequin to the surface world, enabling it to make use of instruments, keep in mind data between steps, and work together with complicated environments.
So, I feel that Allen’s Getting Issues Finished goes to be my place to begin, as that could be a productiveness philosophy versus a group of productiveness instruments. I feel the instruments are endogenous to the philosophy, and never the opposite manner round — at the very least for me. I’ll subsequently in all probability be making an attempt to experiment on right here with sketching out a brand new harness that’s acceptable for the kind of analysis I do. As with abilities, such issues are in all probability extremely idiosyncratic to the person, but in addition as with abilities, I think that there’s an inherent “optimum” construction. I simply don’t assume I do know fairly what it’s but, however I’ve to determine it out if I’m going to concurrently improve my analysis productiveness and cut back errors to zero.
